论文信息

Haodong Liu / Wei Cheng / Minghua Li 等 5 人

Agriculture

论文解读

AI 辅助整理
方法
综述电化学、光学、MEMS传感器与边缘处理器及异构集成,归纳三种设计模式。
发现
土壤监测重选择性校准,表型重采集计算协同,病原检测重样本处理。
机会点
可研究面向真实样本误差与全系统能耗的任务专用硬件评测基准与田间长期验证方法。

论文摘要

农业传感系统必须在能量、响应时间和维护等约束条件下,将多变的土壤、植物和产品信号转化为可靠信息。本综述提出一个任务驱动的框架,将传感器芯片、模拟读出、边缘计算和系统集成贯穿于作物生产与农产品检测之中。内容涵盖电化学与生物化学传感器、微流控与MEMS器件、光学与光谱系统、边缘处理器与加速器,以及从板级到异质集成的各个层面。该综合归纳识别出三种反复出现的设计模式。土壤养分监测高度依赖传感器选择性、参考稳定性和校准。基于图像的植物表型分析需要采集、数据移动与计算的协同配合。在低丰度病原体检测中,在推断性能成为决定性因素之前,样品处理和信号采集可能主导整个工作流程。这些模式引导硬件向主导测量或处理约束的方向进行专门化设计。该综合归纳还将全系统能耗、长期稳定性和田间验证确定为将实验室演示转化为农业应用的重点方向。我们提出针对具体任务的硬件配置和评估标准,涵盖真实样品误差、端到端响应时间、实测系统能耗、数据完整性、校准保持性和生命周期成本。这些成果共同为将硬件选择与集成同农业测量需求相对接提供了实用指导。

全文

Agriculture

Academic Editor: Jibo Yue
Received: 26 October 2026
Revised: 8 November 2026
Accepted: 11 November 2026
Published: 13 November 2026
Copyright: © 2026 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license.
Review
Task-Driven Design and Integration of Sensor Chips and Edge
Computing for Smart Agriculture
Haodong Liu, Wei Cheng, Minghua Li, Chao Feng and Ning Yang *
School of Electrical and Information Engineering, Jiangsu University, Zhenjiang 212013, China;
2112407139@stmail.ujs.edu.cn (H.L.); 2112407121@stmail.ujs.edu.cn (W.C.); 2112507184@stmail.ujs.edu.cn (M.L.);
2112507159@stmail.ujs.edu.cn (C.F.)

  • Correspondence: yangn@ujs.edu.cn
    Abstract
    Agricultural sensing systems must convert variable soil, plant and product signals into
    reliable information under constraints on energy, response time and maintenance. This
    review presents a task-driven framework linking sensor chips, analog readout, edge com-
    puting and system integration across crop production and agricultural product testing. It
    covers electrochemical and biochemical sensors, microfluidic and MEMS devices, optical
    and spectral systems, edge processors and accelerators, and board-level to heterogeneous
    integration. The synthesis identifies three recurring design patterns. Soil nutrient monitor-
    ing depends strongly on sensor selectivity, reference stability and calibration. Image-based
    plant phenotyping requires coordinated acquisition, data movement and computing. In
    low-abundance pathogen detection, sample handling and signal acquisition may domi-
    nate the workflow before inference performance becomes decisive. These patterns guide
    hardware specialization toward the dominant measurement or processing constraint. The
    synthesis also identifies complete-system energy, long-term stability and field validation
    as priorities for translating laboratory demonstrations into agricultural applications. We
    propose task-specific hardware configurations and evaluation criteria covering real-sample
    error, end-to-end response time, measured system energy, data completeness, calibration
    retention and life-cycle cost. Together, these outputs provide practical guidance for aligning
    hardware selection and integration with agricultural measurement needs.
    Keywords: smart agriculture; sensor chips; edge computing; system integration; measurement
    stability
  1. Introduction
    Reliable sensing and timely processing underpin water and nutrient management,
    crop growth and stress monitoring, disease detection, and precision field operations [1].
    Each task defines a distinct hardware objective. Soil nutrient monitoring requires quan-
    titative analyte measurement and stable calibration; plant condition assessment requires
    separation of physiological changes from environmental effects; mobile field operations
    require recognition and decisions within a specified response time [2,3]. These tasks share
    a measurement chain comprising signal acquisition, conversion and readout, processing,
    and output generation. Deploying this chain near soils, plants, and field equipment makes
    size, energy use, environmental robustness, and maintenance central design variables [4].
    Sensor chips and edge computing provide complementary means of tailoring this
    measurement chain. Ion-sensitive field-effect transistors (ISFETs) convert ion-dependent
    Agriculture 2026, 16, 2497 https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 2 of 40
interfacial responses into electrical signals [5,6], while microfluidic chips control sample
transport, reactions, and detection within microchannels [7,8]. Microfabricated optical
structures and integrated photodetectors reduce the size and component count of spectro-
scopic and imaging systems [9]. These sensing technologies can be combined with external
or integrated actuation and readout according to the measurement task. Microcontrollers
(MCUs), embedded processors, and field-programmable gate arrays (FPGAs) provide
local preprocessing, feature extraction, and inference, reducing dependence on continuous
connectivity and cloud processing [10,11]. Together, they support coordinated design of
sensing, readout, and computation within agricultural systems [12,13].
The central design question is where hardware specialization produces a measurable
benefit across the agricultural workflow. For sensing, sample matrices, environmental
fluctuations, and drift determine how laboratory detection performance transfers to real
measurements [14]. For edge computing, memory capacity , bandwidth, operator compatibility ,
and power determine how efficiently a model executes on the target platform [15]. Integration
then couples these choices through power delivery , heat transfer, electrical interfaces, and
environmental protection [16]. A workflow-based assessment links each hardware advance to
its contribution to sample handling, signal quality , processing, or sustained operation. It also
identifies the external components needed to deliver that contribution.
Existing reviews of agricultural sensing and edge intelligence offer valuable insights
into sensing modalities, computational methods and hardware implementations [17–19].
However, gains at the device or algorithm level may not translate into improved perfor-
mance across a complete agricultural measurement system. This review connects these
perspectives through a task-driven engineering framework that links agricultural require-
ments to hardware design. We examine how these requirements guide the selection and
coordination of sensing interfaces, analog readout circuits, computing resources and integra-
tion architectures. We distinguish experimentally validated agricultural implementations
from transferable hardware demonstrations and translate their reported capabilities and
limitations into application-specific design choices. By linking dominant system bottlenecks
to hardware configurations, deployment constraints and evaluation criteria, the framework
supports practical engineering decisions. It clarifies when dedicated hardware is warranted
and when modular, general-purpose solutions are more appropriate.
Section Literature Scope and Mapping Overview defines the literature scope and
mapping approach. Section 2 translates agricultural tasks into hardware requirements.
Section 3 examines sensing mechanisms and readout architectures, and Section 4 con-
nects edge workloads to model compression, hardware acceleration, and dataflow design.
Section 5 compares integration approaches, while Section 6 addresses measurement sta-
bility, protection, and maintenance. Section 7 develops priorities for sensing–computing
co-design, module reuse, and comparable evaluation. Section 8 summarizes the resulting
design principles.
Literature Scope and Mapping Overview
This topic-oriented narrative review examines sensing, readout, computing, and sys-
tem integration technologies for agricultural applications. Bibliographic searches were
conducted in the Web of Science Core Collection, covering 2000–2026 with a cutoff date of 1
October 2026. Two topic-field (TS) search strategies were employed for different analytical
purposes. A broad search combining agricultural terms with integrated circuits and chip-
related technologies retrieved 15,033 records for the annual publication analysis in Figure 1a.
A focused search combining agricultural applications with sensing and computing hardware
terms yielded 2280 candidate records for the keyword co-occurrence analysis in Figure 1b.
The complete Boolean search expressions are provided in Supplementary Table S1.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 3 of 40
Figure 1. Publication trends, keyword mapping, and review framework for sensing and computing
research in smart agriculture. (a) Annual publication counts from a broad Web of Science Core
Collection search (n = 15,033). (b) VOSviewer keyword co-occurrence network based on 1886 records
retained after screening and retraction removal. (c) Conceptual framework linking agricultural
requirements to sensing, computing, and system integration. Searches covered 2000–2026 (cutoff:
1 October 2026); 2026 represents an incomplete year.
The 2280 records from the focused search were exported with full records and cited
references. WoS accession numbers and available DOIs were checked for duplication, and
titles and abstracts were manually screened. Original studies and reviews addressing
agricultural sensing, signal readout, computing hardware, system integration, or relevant
transferable hardware principles were considered for inclusion. Records outside the review
scope or retrieved through incidental or ambiguous keyword matches were excluded,
while supporting background literature was classified separately. Screening identified
1894 included records, 367 excluded records, and 19 background records. Eight retracted
publications were subsequently removed from the included set, leaving 1886 records for
keyword co-occurrence analysis. Keyword variants, abbreviations, and synonyms were
standardized using a thesaurus before visualization in VOSviewer 1.6.20.
The three panels of Figure 1 provide complementary perspectives on the research land-
scape and the framework adopted in this review. Figure 1a illustrates publication trends,
while Figure 1b identifies thematic relationships within the curated literature. Building
on these perspectives, Figure 1c presents a task-driven conceptual framework that links
agricultural requirements, including weak signals and biological variability, multimodal
data, field conditions, and resource constraints, to three interconnected hardware design
dimensions: dedicated sensing and readout, computing and storage co-design, and hetero-
geneous integration and packaging. This framework guides the subsequent comparison
of sensing technologies, computing architectures, integration strategies, and deployment
constraints. The broader publication-trend dataset and the curated keyword-mapping
dataset serve distinct descriptive purposes, while the technical synthesis additionally con-
siders representative hardware studies from related engineering fields. Data for 2026 reflect
publications available up to the cutoff date rather than a complete calendar year.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 4 of 40
2. Agricultural Tasks and Hardware Design Requirements
Agricultural tasks define hardware requirements through their measurement targets,
response times, and operating conditions. Soil ion monitoring prioritizes selectivity, sta-
ble references, and control of environmental interference. Field image recognition links
acquisition quality to processing throughput and response latency. Distributed monitoring
adds energy, communication, and maintenance budgets. Figure 2 illustrates these require-
ments through plant-mounted sensors, UA V-based multimodal observation, distributed
field nodes, and wireless networks. The following subsections connect weak signals and bio-
logical variability, multimodal data, power and environmental conditions, and deployment
costs to specific sensing, computing, and integration choices.
Figure 2. Representative implementations of sensing and field monitoring systems in smart agri-
culture. (a) Plant physiological monitoring using leaf-mounted flexible electronic electrodes [20];
(b) UA V-based multimodal observation system integrating hyperspectral imaging, visible-light imag-
ing, thermal infrared imaging, and light detection and ranging (LiDAR) [21]; (c) solar-powered
field monitoring node integrating soil moisture sensing and wireless communication [22]; (d) in
situ monitoring system comprising buried soil sensors and an aboveground wireless communi-
cation module [23]; (e) LoRaWAN-based distributed monitoring and centralized management
architecture for vineyard and greenhouse environments [24]. All panels were adapted from
the cited sources under the Creative Commons Attribution 4.0 International License (CC BY 4.0;
https://creativecommons.org/licenses/by/4.0/ (accessed on 25 October 2026)). Typography and
layout were standardized without altering the original technical content.
2.1. Weak Signals and Biological Variability
Hardware for agricultural sensing must separate target-related signals from biological
and environmental variability. Soil ion activities change with moisture, salinity, organic
matter, and temperature [25]. Plant metabolite concentrations vary across tissues and
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 5 of 40
over the day [26], while leaf reflectance and thermal emission depend on cultivar, growth
stage, illumination, orientation, and background [27]. Sensor installation and changes
at the sensing interface add further variation. Reliable acquisition therefore combines
adequate sensitivity with reference measurements, stable interfaces, and control of the
dominant disturbances.
Weak-signal acquisition depends on matching the sensing interface to its conditioning
and readout circuit. Potentiometric sensors need high input impedance, low input bias
current, and a stable reference to preserve the measured potential [28]. Amperometric
sensors require control of leakage, background current, and circuit noise [29]. Optical
systems balance exposure, detector dynamic range, and signal-to-noise ratio to retain
information across weak and strong responses [30]. Repeated measurements and averaging
suppress random noise; systematic drift, nonspecific responses, and sample-dependent
bias require targeted correction or interface design. Identifying the dominant error source
directs improvements to the sensing, conversion, or readout stage where they have the
greatest effect.
Specificity and traceability support interpretation across variable biological samples.
Because a spectral or electrical feature can respond to several physiological processes and
environmental factors, control experiments and independent reference measurements estab-
lish its association with the target state [31]. Cultivar, growth stage, and sampling location
can shift baselines and response characteristics. Recording acquisition time, temperature,
moisture conditions, and calibration status provides context for these shifts and supports
signal correction, assessment of model applicability, and validation across conditions.
2.2. Multimodal Data and Real-Time Processing
Multimodal agricultural sensing combines visible-light, thermal infrared, and spectral
observations with environmental measurements such as temperature and humidity [32].
Each modality contributes a different view of the target and imposes distinct sampling,
spatial resolution, and data-volume requirements. Useful combinations require temporal
synchronization, spatial registration, and scale alignment appropriate to the task. Com-
paring individual and combined modalities establishes the information gained from each
channel. This evidence guides channel selection, interface and buffering requirements,
cross-sensor calibration, and the stage at which information is fused.
The target’s dynamics and the response time of the agricultural operation define real-
time requirements. Soil moisture, nutrient levels, and environmental conditions can often be
monitored through intermittent sampling and periodic processing [33]. Mobile operations
require recognition and localization results to remain aligned with acquisition timestamps
and equipment positions [34]. Depending on the measurement principle, the response
interval includes exposure or scanning, sensor stabilization, sample transport and reaction,
readout, transmission, preprocessing, inference, and output generation [35]. Measuring
these stages identifies the dominant delay and establishes whether the complete workflow
meets the task deadline.
Acquisition and processing schedules can be matched to both target dynamics and
available energy. Continuous image streams benefit from pipelining, local buffering, and
data reuse that balance acquisition rate, computing throughput, and transmission band-
width [36]. Slowly varying soil and environmental measurements support periodic sam-
pling, sleep modes, and event-triggered operation. For event-triggered systems, thresholds,
wake-up overhead, and missed-event rates determine the useful operating range. Evaluat-
ing the complete cycle connects energy savings to the required observation interval and
response time.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 6 of 40
2.3. Power Supply and Environmental Constraints
Energy budgeting determines the operating schedule of distributed agricultural nodes.
Battery- and solar-powered systems draw energy across sensing, readout, processing, trans-
mission, and standby states [37]. Optical illumination and fluid actuation add task-specific
loads. Their contributions vary with sampling frequency, data volume, communication
protocol, and duty cycle, so the benefit of a low-power processor depends on the share
of energy used for digital processing [38]. Solar-powered nodes also coordinate variable
harvesting with storage capacity [39]. Component selection, sampling, transmission, and
sleep schedules should therefore be designed together to sustain the required observations
within the available energy budget.
Environmental design requirements follow from the deployment location and the
measurement pathway [40]. Temperature and humidity fluctuations, condensation, dust,
and chemical exposure affect sensing interfaces, optical windows, connections, and elec-
tronic components [41]. Buried sensors require stable contact with soil and protection
against contamination. Plant-mounted devices must accommodate growth, deformation,
and changes in adhesion. Machinery-mounted modules face vibration, supply fluctuations,
and deposited material [42]. Enclosed electronics also require a balance between sealing and
heat dissipation. Separating exposed sensing interfaces from protected readout, computing,
and interconnection regions allows protection to be tailored to each function.
Environmental protection is most effective when incorporated into device selection and
architecture. Soil-contact, plant-contact, and optical sensors need accessible measurement
pathways, while their readout circuits, power modules, and data connections require
appropriate placement and protection. Partitioned packaging, replaceable sensing elements,
and accessible calibration interfaces can support this arrangement. The sensing principle,
deployment duration, and available field maintenance determine the suitable configuration.
2.4. Application Diversity and Cost Constraints
Reusable electronics can accommodate agricultural diversity when task-specific func-
tions are clearly identified. Established sensors and general-purpose modules support
routine temperature, humidity, and soil moisture measurements. Specific compounds,
physiological indicators, and visual targets may require tailored recognition interfaces,
sampling structures, or processing methods [43]. Signal conditioning, conversion, control,
and communication offer opportunities for reuse, while sensing elements and specialized
processing remain closely tied to the target. This partitioning concentrates development
effort on the functions that determine measurement performance.
Life-cycle cost provides the economic basis for choosing an agricultural hardware
architecture. Relevant costs include development, validation, procurement, manufacturing,
assembly, calibration, and maintenance [44]. General-purpose boards and modular systems
offer short development cycles and flexibility for prototypes and small deployments with
evolving requirements; their size, power, and interface overhead become important as
deployment scales [45]. Application-specific integrated circuits (ASICs) tailor circuitry and
resources to stable functions and may reduce unit cost at sufficient volume, with upfront
investment in design, verification, and fabrication [46]. The choice therefore follows from
requirement stability, anticipated scale, the dominant performance bottleneck, and total
life-cycle cost.
Translating these requirements into hardware choices begins with identifying the dom-
inant constraint in the target agricultural workflow. When weak signals, matrix interference
or sensor drift dominate measurement uncertainty, priority should be given to the sensing
interface, analog readout and calibration. When data volume or acquisition frequency
drives resource consumption, channel selection, buffering and dataflow optimization be-
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 7 of 40
come central design considerations. For computationally intensive tasks with strict response
deadlines, embedded accelerators or FPGA pipelines may be warranted after assessing soft-
ware and model-level optimization. Integration choices should balance miniaturization with
power requirements, environmental exposure, interface complexity, maintainability and
life-cycle cost. This task-to-hardware reasoning underpins the representative configuration
matrix in Section 7.1.
3. Sensing Technologies and Readout Architectures
Sensing architectures link target recognition to signal transduction and readout for
downstream processing. This section examines electrochemical and biochemical sensors, mi-
crofluidic and microelectromechanical systems (MEMS), optical spectroscopy and imaging,
and multimodal sensing architectures. Figure 3 illustrates representative sensing devices
and readout systems, while Table 1 summarizes their applications, performance and inte-
gration configurations. Analog front-end architectures receive particular attention because
they shape signal conditioning, conversion accuracy and compatibility with downstream
computing hardware. By linking sensing mechanisms to circuit-level requirements, this
section identifies integration choices suited to agricultural tasks and deployment conditions.

Figure 3. Representative technologies for agricultural sensor chips and integrated microsystems.
(a) Graphene field-effect transistor gas sensor array and its experimental detection platform [47];
(b) microfluidic impedance sensing system for detecting rice-associated particles, comprising a
microfluidic chip, miniature air pump, and external impedance analyzer [48]; (c) integrated miniature
spectrometer chip and its optical dispersive structure [49]; (d) lensless diffraction imaging system
based on LED illumination and a CMOS image sensor [50]; (e) microfabrication process of a flexible
microstructured sensor and the associated readout system architecture incorporating capacitive
sensing, signal conditioning, analog-to-digital conversion, and wireless communication [51]. All
panels were adapted from the cited sources under the Creative Commons Attribution 4.0 International
License (CC BY 4.0; https://creativecommons.org/licenses/by/4.0/ (accessed on 25 October 2026)).
Typography and layout were standardized without altering the original technical content.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 8 of 40
Table 1. Representative agricultural sensor chips and integrated microsystems.
Application Device and Sensing Principle System Configuration and External Dependencies Reported Performance and Validation Scope
Soil nutrients
Si-based ChemFET/ISFET; ion-selective
potentiometry [52]
Si FETs, Pt electrodes, and PCB packaging;
independent system operation not established
Greater than 50 mV/decade; 15-day soil monitoring;
stability affected by soil contact
Photocurable membrane ISFET; pH, Ca2+ ,
K+ , and NO3 − detection [53]
ISFETs with selective polymer membranes;
complete autonomous readout not established
Validated in solutions and horticultural soils against
reference methods
Ionophore-based ChemFET; NH4 + /NO3 −
potentiometry [54]
Selective membranes integrated with ChemFETs;
standalone system operation not established
Approximately 56 mV/pX; liquid and soil validation;
nitrate-sensor drift observed
rGO-based FET; NH4 + /NO3 − current
response [55]
Ti/Au–rGO sensing device with external biasing
and electrical analysis
0.9/0.45 μA/ppm sensitivity; 35/41 s response time;
externally instrumented evaluation
Soil properties
Shielded impedance chip;
moisture-dependent impedance [56]
On-chip shielding and leakage-current separation;
complete field node not demonstrated
Detection of low water content (≤10%);
slope-monitoring application rather than direct
agricultural field validation
PMMA capacitive moisture sensor [57]
Capacitive device with PCB readout and GSM
communication; independent power and
maintenance requirements not established
Four electrode designs; calibration and temperature
effects evaluated
3D-printed Ni capacitive sensor [58] 60 MHz capacitive measurement with an ESP32-based modular system Volumetric water content 0–0.40 cm3 /cm3 ; three soil textures; laboratory validation
Plant physiology
Antibody-functionalized SPR chip;
abscisic acid detection [59]
CM5 sensing chip measured using a benchtop
Biacore 8K instrument
LOD 1.36 ng/mL; tomato xylem samples validated; not
a standalone portable reader
Biodegradable GOx microneedle; glucose
detection [60]
Conductive hydrogel sensing interface; complete
three-electrode and wireless readout integration
remains to be demonstrated
Approximately 14 h of leaf monitoring; temperature
and humidity correction
HRP/rGO microneedle array; H2 O2
amperometry [61]
Integrated three-electrode array and flexible PCB;
external electrochemical workstation
LOD 0.06 μM; approximately 1 min response;
validated in tobacco and soybean
Crop pathogens
PDMS–glass microfluidic chip; spore
enrichment and optical attenuation [62]
External pump, light source, optical fibers, and
spectrometer required
Greater than 80% enrichment; externally actuated and
instrumented laboratory system
QCM acoustic biosensor; PCR-product
binding [63]
Separate PCR procedure and commercial
QCM reader
Three bacterial pathogen targets differentiated using
PCR amplicons of 288, 422, and 689 bp; external PCR
and commercial QCM readout required
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 9 of 40
Table 1. Cont.
Application Device and Sensing Principle System Configuration and External Dependencies Reported Performance and Validation Scope
Crop pathogens CMOS micro-GC system; VOC separation and thermal-conductivity detection [64] CMOS detector, GC column, and Bluetooth communication; independent sample-to-result autonomy not established Wireless VOC measurement for indirect pathogen-related sensing
Pesticides
Aptamer-functionalized carbon FET [65]
Magnetic nanoporous carbon–FET interface;
regeneration required; complete readout
configuration not specified in the comparison
Up to 40 sensing-material reuse cycles; full-system
autonomy not established
Graphene FET immunosensor;
chlorpyrifos sensing [66]
Wire-bonded sensor with external source-meter
and lock-in-amplifier readout
Chlorpyrifos detection demonstrated with external
electrical instrumentation
Microfluidic Au-electrode immunochip;
impedance spectroscopy [67]
Microchamber, functionalized electrodes,
approximately 40 min incubation, and external
impedance analyzer
LOD 0.5 ng/mL; 75.2–96.5% recovery in vegetable
samples
pSBMA paper-based chip; colorimetric
detection [68]
Paper/PDMS–glass reaction chip; external imaging
and ImageJ-based quantification
LODs: 0.235 mg/L (chlorpyrifos), 4.891 mg/L
(profenofos), and 4.053 mg/L (cypermethrin);
externally imaged colorimetric assay
Food safety
SAM-functionalized SPR chip; mycotoxin
detection [69]
SPR chip with benchtop reader; HPLC–MS/MS
used for reference comparison
LODs: 0.59, 1.27, 7.07, and 3.26 ng/mL for AFB1, OTA,
ZEN, and DON, respectively; corn/wheat testing and
HPLC–MS/MS comparison; benchtop reader
Imaging SPR microarray; competitive
immunoassay [70]
Microarray preparation, fluid handling, imaging,
and iSPR reader require external equipment
DON LODs: 21/17 ng/mL in maize/wheat extracts;
ZEN LODs: 16/10 ng/mL; multiplex assay with
external fluidic and imaging equipment
USB-compatible immunochip; AFB1
impedance immunoassay [71]
Printed electrodes coupled to a portable detector;
independent power and automated sample
preparation not established
AFB1 LOD of 5 ng/mL; total detection time below 1 h
including incubation; portable reader tested with rice
samples, but autonomous sample preparation not
demonstrated
OWLS sensor chip; DON evanescent-field
sensing [72]
Optical waveguide sensing with flow injection and
associated optical readout
Label-free DON measurements; direct-assay sensitivity
inadequate for the tested grain samples
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 10 of 40
In evaluating the reported sensing technologies, it is important to distinguish device-
level integration from the capabilities of a complete measurement system. A sensing chip
may perform target recognition, enrichment, or signal transduction while relying on ex-
ternal pumps, optical instruments, electrochemical workstations, or sample preparation
procedures. Similarly, portable readout and local edge processing do not necessarily imply
autonomous operation. Table 1 therefore identifies the external equipment and processing
dependencies of representative implementations alongside their reported measurement
performance. System autonomy is considered demonstrated only when the intended mea-
surement workflow, including necessary sample handling, actuation, readout, processing,
and result generation, has been validated without reliance on external laboratory instru-
mentation or step-by-step manual intervention. Where such evidence is unavailable, the
demonstrated functionality is described without assuming complete system autonomy.
3.1. Electrochemical and Biochemical Sensing
Electrochemical and biochemical sensors provide selective access to soil nutrients,
plant physiological indicators, and pesticide residues. Electrochemical transducers mea-
sure changes in interfacial potential, current, or impedance, while enzymes, antibodies,
and aptamers add target recognition [73]. ISFETs translate sensing-interface potentials
into transistor outputs that can be coupled to microelectronic readout. Electrode-based
sensors capture responses from interfacial reactions or molecular binding [74]. Reference
conditions, bias stability, signal conditioning, and noise determine how these responses
become quantitative measurements. Coordinating the sensing element and readout circuit
therefore supports both measurement quality and miniaturization.
Transistor-based soil probes illustrate how selective interfaces can support multianalyte
monitoring. Packaged pH-ChemFETs and NH4 + - and NO3 − -selective ISFETs enabled con-
tinuous soil measurements over 15 days; the study also identified moisture and soil contact
as determinants of stability [52]. Photocurable ion-selective membranes extended ISFET
measurements to pH, Ca2+ , K+ , and NO3 − , with validation against reference methods in so-
lutions and horticultural soils [53]. Fluoropolysiloxane membrane-based ChemFET/ISFET
devices measured NH4 + and NO3 − in liquid and soil samples, with membrane-dependent
stability and substantial nitrate-sensor drift [54]. Together, these results connect ion selec-
tivity to the reference, contact, and calibration requirements of soil monitoring and guide
the design of associated readout modules.
Microneedles provide direct interfaces for monitoring plant biochemical signals.
Biodegradable conductive hydrogel microneedles incorporating glucose oxidase (GOx) de-
tected glucose through enzymatically generated H2 O2 and supported approximately 14 h of
leaf monitoring with temperature and humidity correction. Complete three-electrode inte-
gration and wireless readout remain development steps [60]. An array functionalized with
horseradish peroxidase (HRP), chitosan, and reduced graphene oxide (rGO) integrated
working, reference, and counter electrodes. It achieved an H2 O2 detection limit of 0.06
μM in standard solutions and validation in tobacco and soybean using an external electro-
chemical workstation [61]. These implementations connect tissue-compatible recognition
to multielectrode sensing. Compact readout requires low-noise current measurement and
stable biasing, with tissue contact, enzyme activity, and matrix effects guiding calibration.
Biological recognition coupled to transistors or microelectrodes enables compact, se-
lective pesticide measurements. A field-effect transistor (FET) biosensor using bridge-type
aptamers and magnetic nanoporous carbon supported up to 40 regeneration cycles of the
sensing material [65]. An antibody-functionalized graphene FET detected chlorpyrifos
through changes in channel electrical properties, with source meters and lock-in amplifiers
providing readout [66]. A microfluidic immunosensor with antibody-functionalized gold
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 11 of 40
interdigitated microelectrodes achieved a chlorpyrifos detection limit of 0.5 ng/mL and
vegetable-sample recoveries of 75.2–96.5%, using approximately 40 min of incubation and
an external impedance analyzer [67]. These studies establish recognition and transduc-
tion capabilities at the device level. Their workflows identify integration targets: sample
preparation, regeneration or incubation, compact readout, and quantitative validation in
real samples.
3.2. Microfluidic and MEMS-Based Sensing
Microfluidic and MEMS technologies tailor sample handling and signal acquisition
through microscale structures. Microfluidic channels control transport, mixing, enrichment,
and reactions, and can incorporate optical, electrochemical, or impedance detection [75].
MEMS structures sense or actuate mechanical, thermal, and fluid-related quantities, with
readout matched to the transduction mechanism. The location of each function defines
the system architecture: sample processing and reactions can reside on a microfluidic
chip while optical or electrical instruments perform readout, and MEMS elements can
connect to off-chip conditioning and processing [76]. This functional description makes
their contributions to integration explicit.
Controlled transport and enrichment give microfluidics a distinct role in agricultural
particle and chemical analysis. A PDMS–glass chip enriched airborne spores with efficiency
exceeding 80% and estimated their concentration from transmitted-light attenuation. An
external pump, light source, optical fibers and spectrometer supported sample actuation
and optical measurement [62]. A multireaction chip combined functionalized filter paper
with patterned PDMS–glass structures for parallel colorimetric detection of chlorpyrifos,
profenofos and cypermethrin, with external imaging and quantification [68]. Together, these
implementations demonstrate on-chip enrichment and parallel reaction processing within
instrument-supported detection workflows. Further system integration can coordinate
these assay functions with compact actuation and optical instrumentation, data processing
and power management.
Microstructural design links miniaturization to repeatable sample transport and stable
measurement. Channel geometry, wettability, and sealing govern microfluidic flow and
reaction consistency; particulate agricultural samples require dimensions and actuation
that accommodate flow resistance and clogging. Silicon interdigitated electrodes coated
with Ti3 C2 Tx MXene measured soil moisture across gravimetric water contents of 1–18%,
with texture-dependent responses and hysteresis identifying the need for soil-specific cali-
bration [77]. A separate MEMS study coordinated temperature and pressure sensing with a
backside pressure-inlet package. The package passed a 30 MPa test, and the sensor delivered
5.536 mV/MPa sensitivity over 0–15 MPa with temperature error below ±0.61 °C [78]. This
non-agricultural demonstration provides a packaging design example. Applying such ap-
proaches to agricultural systems calls for repeatable fabrication, real-sample measurements,
and validation of actuation, readout, and stability under the intended conditions.
3.3. Optical, Spectroscopic, and Imaging Chips
Optical sensing captures morphological, compositional, and physiological information
through reflection, absorption, scattering, fluorescence, and thermal emission. Image sen-
sors emphasize spatial resolution, dynamic range, and signal-to-noise ratio. Spectroscopic
devices balance spectral selectivity, throughput, and calibration stability, while surface
plasmon resonance (SPR) and waveguide sensors combine optical coupling with surface
recognition [79,80]. Compact layouts, dispersive structures, filter arrays, waveguides, and
integrated photodetectors provide several routes to miniaturization [81]. The resulting
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 12 of 40
system configuration is defined by how these elements connect to illumination, scanning,
and readout.
Integrated wavelength selection and photodetection reduce the component count of
spectroscopic systems. Zheng et al. placed Au/SiNx plasmonic narrowband filter arrays
directly on InGaAs photodetectors to form a 16-channel near-infrared spectrometer covering
900–1700 nm. The filters had approximately 5 nm full width at half maximum; computa-
tional reconstruction separated simulated peaks 1 nm apart near selected wavelengths [82].
Venditti et al. coupled a piezoelectric resonator to a spectrally selective infrared metasurface,
producing a MEMS detector with a 2.3 ms thermal time constant. An infrared source, mul-
tipass gas cell, and low-noise readout completed a compact trace-gas sensing system [83].
These examples demonstrate computational multichannel spectroscopy and selective in-
frared detection, respectively, under their reported non-agricultural conditions [82,83].
Their potential agricultural use requires separate validation with relevant sample matrices,
including assessments of spectral response, calibration stability, and the contribution of
external optical and readout components.
Functionalized optical surfaces support label-free and multiplexed detection. A self-
assembled monolayer (SAM)-modified SPR chip measured four mycotoxins with a benchtop
reader and comparison against high-performance liquid chromatography–tandem mass
spectrometry (HPLC–MS/MS) [69]. Toxin–protein microarrays prepared by continuous-
flow microspotting enabled parallel mycotoxin detection through CCD/CMOS-based imag-
ing SPR (iSPR), with separate microarray preparation, fluid handling, and imaging equip-
ment [70]. Optical waveguide lightmode spectroscopy (OWLS) was evaluated for deoxyni-
valenol (DON); direct-assay sensitivity and response stability were inadequate across the
tested grain samples [72]. These results establish surface recognition and spatial multiplex-
ing capabilities, with assay format, matrix response, and reader configuration determining
the measurement range and deployment format.
Computational optics exchanges part of the optical hardware complexity for calibrated
digital processing. Lensless diffraction imaging records sample-dependent patterns for
direct analysis or reconstruction, while computational spectroscopy recovers spectral infor-
mation from encoded multichannel responses [84]. Optical modulation, detector sampling,
and reconstruction must be designed together. Illumination, sample-to-detector distance,
and optical geometry define the calibration conditions, and the reconstruction workload
determines memory, latency, and energy requirements. Evaluating these quantities together
establishes the value of compact optics for a specific agricultural task.
3.4. Multimodal Readout and Sensing Integration
Multimodal sensing is most useful when its channels resolve complementary aspects
of an agricultural target. Environmental measurements contextualize electrochemical or
spectral responses, and imaging supplies spatial and morphological information [85]. Com-
parisons of individual and combined modalities, including channel ablation, establish each
channel’s contribution [86]. Excitation, signal range, sampling frequency, and calibration
then determine interface design, alignment, and interference control. Selecting channels on
this basis connects information gain directly to circuit and system requirements.
Multichannel operation can align measurement frequency with energy and mainte-
nance needs. Low-power channels can run periodically, while energy-intensive channels
activate on a schedule or in response to events. Warm-up, stabilization, recalibration, and
target dynamics determine the appropriate intervals and thresholds. Because exposed
sensing interfaces may age faster than electronics, replaceable sensing elements paired
with reusable readout modules can simplify maintenance. Electrical compatibility, repeat-
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 13 of 40
able connection, and calibration retention allow this modular arrangement to support
measurement continuity over the deployment period.
3.5. Analog Front-End Architectures and Sensor–Readout Integration
Analog front-end (AFE) circuits connect sensing elements to digital processing units
through signal conditioning, excitation, amplification and conversion. Sensor electrical
characteristics guide the choice of AFE architecture. Potentiometric sensors require high-
input-impedance voltage readout and stable reference potentials. Amperometric sensors
typically use potentiostat circuits to maintain electrode bias and transimpedance ampli-
fiers (TIAs) to convert small currents into measurable voltages. Impedimetric sensors
require controlled AC excitation and phase-sensitive measurement, whereas capacitive
sensors use charge-based readout or capacitance-to-digital conversion. Optical detectors
place additional demands on noise suppression, dynamic range and sampling bandwidth.
Table 2 summarizes the corresponding AFE architectures, critical electrical parameters and
integration considerations.
Table 2. Comparison of representative analog front-end architectures and their integration considera-
tions for agricultural sensing.
Sensing Modality Typical AFE Architecture Key Electrical Parameters Integration Approach Main Constraints
Potentiometric
(ISE/ISFET)
High-impedance voltage
buffer; reference and
bias control
Input impedance, bias
current, offset voltage,
drift, reference stability
Discrete AFE or
integrated readout with
ADC and MCU
Leakage, reference drift,
soil-contact variation
Amperometric Potentiostat, TIA, low-pass filter, ADC Current noise, TIA gain, bias accuracy, dynamic range, bandwidth Integrated potentiostat–TIA or dedicated electrochemical AFE Leakage, electrode polarization, interference
Impedimetric
AC excitation,
current/voltage readout,
synchronous
demodulation or DFT
Frequency range,
amplitude and phase
accuracy, excitation
level, dynamic range
Impedance-converter
IC or integrated
excitation and readout
Parasitic impedance,
electrode interface,
frequency-dependent
calibration
Capacitive
Charge amplifier,
switched-capacitor
circuit, or CDC
Capacitance resolution,
parasitic capacitance,
drift, sampling rate
CDC with MCU or
integrated
MEMS–CMOS readout
Cable parasitics,
humidity, shielding
Optical
(photodetector)
Photodiode, TIA,
low-pass filter, ADC
Responsivity, dark
current, noise, dynamic
range, bandwidth
Integrated
photodetector–TIA
readout with ADC and
MCU/SoC
Ambient light
interference, dark-current
drift, optical alignment,
temperature sensitivity
AFE design shapes measurement fidelity and chip-level integration. In potentiometric
interfaces, low input leakage and reference stability are central to reliable long-term soil
measurements. For amperometric readout, TIA feedback resistance and amplifier noise
help define the detectable current range and measurement bandwidth. Impedance measure-
ments require coordinated excitation and detection, with phase error, electrode polarization
and parasitic capacitance shaping performance across the frequency sweep. Capacitive
and optical readout benefit from shielding, baseline correction and controlled integration
or sampling times to preserve weak signals under changing environmental conditions.
Evaluating these sensor-specific requirements alongside ADC resolution, sampling rate,
power consumption and calibration stability establishes the electrical design priorities for
each application.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 14 of 40
Representative integrated AFEs illustrate the trade-offs between measurement func-
tionality and circuit complexity. For example, the LMP91000 provides programmable
transimpedance gain from 2.75 to 350 kΩ with supply current below 10 μA under specified
operating conditions [87], whereas the AD5940 integrates electrochemical excitation, cur-
rent readout, a 16-bit ADC, and impedance measurement capabilities up to 200 kHz [88].
For capacitive sensing, the FDC1004 provides four measurement channels with a specified
resolution of 0.5 fF and integrated shielding [89,90]. These specifications illustrate possible
integration approaches but do not establish measurement accuracy or field reliability in
agricultural environments. Such performance must be assessed after connecting the actual
sensing element and accounting for environmental interference, calibration, and peripheral
power consumption.
AFE integration ranges from discrete conditioning circuits and dedicated readout ICs
to mixed-signal system-on-chip architectures. Integrated readout can shorten interconnects
and reduce external component count, while shared amplifiers, multiplexers and ADCs
can improve area and power efficiency in multichannel systems. Effective resource sharing
requires coordinated switching and acquisition timing to manage settling, switching tran-
sients and cross-channel interference. Silverio et al. demonstrated a multiparameter readout
chip supporting potentiometric, amperometric and impedimetric sensing, with measures to
address interference between measurement modes [29]. For agricultural deployment, mois-
ture and contaminant exposure, sensor replacement needs and calibration access further
guide the integration strategy. When sensors have shorter service lives than processing hard-
ware, separating exposed sensing interfaces from protected readout electronics supports
independent sensor replacement and maintenance.
4. Agricultural Edge Computing and Hardware Acceleration
Agricultural edge computing enables timely local decisions by performing preprocess-
ing, feature extraction and inference on sensor nodes or nearby devices. Workload charac-
teristics guide the choice of computing platform, model compression strategy, dedicated
processing hardware and dataflow architecture. Figure 4 illustrates these implementation
routes. Table 3 compares representative implementations across input specifications, task
performance, latency or throughput, power consumption, memory requirements and vali-
dation conditions. The comparison covers agricultural prototypes evaluated in laboratory
or field settings alongside hardware demonstrations developed outside agriculture with
potential for agricultural adaptation. Reported performance is interpreted in the context of
measurement boundaries and external resource requirements to clarify the deployment
conditions under which each implementation offers practical value.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 15 of 40

Figure 4. Edge computing and hardware acceleration architectures for agriculture. (a) Machine learn-
ing workflow for resource-constrained devices, from preprocessing, training, and validation to hard-
ware adaptation, testing, and runtime monitoring [91]; (b) ARM–FPGA crop row detection using AXI
interconnects, video direct memory access (VDMA), segmentation, and row detection modules [92];
(c) plant disease recognition combining inter-class and intra-class knowledge distillation (top) with
an ARM–FPGA neural network accelerator (bottom) [93]; (d) Zynq-7035 spectral processing system
and prototype, with processor–programmable logic partitioning, preprocessing, reflectance calcula-
tion, and spectral analysis [94]; (e) Nebula multibranch neural network accelerator with processing
element (PE) clusters, routers, on-chip SRAM, and interconnects for feature-map and parameter
transfer and reuse [95]. All panels were adapted from the cited sources under the Creative Commons
Attribution 4.0 International License (CC BY 4.0; https://creativecommons.org/licenses/by/4.0/ (ac-
cessed on 25 October 2026)). Typography and layout were standardized without altering the original
technical content.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 16 of 40
Table 3. Comparison of agricultural edge computing implementations and transferable hardware demonstrations under reported evaluation conditions.
Application and
Platform Input Specification Task Performance Reported Timing Power Memory and Resources Validation Scope and Limitations
Strawberry ripeness
classification; Cyclone
IV/MAX 10 FPGA +
MCU [96]
Our ripeness classes;
image resolution NR F1 = 0.933
533 ms per classification
pipeline; 495 ms for
feature extraction
NR External memory; footprint NR
Agricultural laboratory
prototype; acquisition and
system energy not
characterized
Crop yield classification;
PYNQ-Z2
ARM–FPGA [97]
Tabular agricultural data
96.43% execution-time
reduction relative
to CPU
Relative execution time
reported; complete
workflow latency NR
1.657 W, reported
on-chip power
BRAM use reported;
runtime memory
footprint NR
Agricultural-data hardware
demonstration; field operation
not established
Agricultural video
dehazing; XC7S25
FPGA [98]
1920 × 1080 video Real-time dehazing
60 FPS processing
throughput; complete
decision latency NR
Less than 2 W,
reported FPGA
power
LUT/FF use reported;
memory footprint NR
Agricultural imagery
evaluated; downstream
recognition and sustained
deployment not established
Agricultural field image
monitoring; Cyclone IV
FPGA [99]
800 × 600 RGB images CCC = 0.99/0.91 in laboratory/outdoor tests 40 FPS image acquisition and processing NR Memory footprint NR Laboratory and outdoor prototype validation; external PC and power equipment used
Weed detection; Jetson
Nano [100]
640 × 640; FP16;
batch size 1 mAP@50 = 87.14% 7.2 FPS edge inference; full pipeline latency NR NR 2.63 M model parameters; runtime footprint NR Agricultural dataset and edge-board evaluation; field operation not established
Irrigation management;
Raspberry Pi [101]
Soil and weather
sensor data
Soil moisture prediction
and irrigation control
Edge service latency
313.6 ms; response time
393.8 ms
NR
Approximately 31%
RAM utilization;
absolute footprint NR
Agricultural IoT edge-service
implementation; long-term
deployment performance
not established
Mushroom cultivation
monitoring; XIAO
ESP32-S3 [102]
Environmental sensor
measurements
Local TinyML-based
monitoring and control Inference latency NR NR
8 MB PSRAM and 8 MB
flash available;
application footprint NR
Agricultural
cultivation-environment
evaluation; energy and
long-term reliability
not quantified
Ultra-low-light in-sensor
computing; SNSPD
array [103]
25-pixel
photon-counting array
92.22% accuracy at 0.12
photons/pixel/pattern
Complete system latency
NR
Cooling-system
power NR NR
Non-agricultural cryogenic
demonstration; approximately
2.3 K operation
Notes: NR, not reported or not established for the specified evaluation; CCC, Lin’s concordance correlation coefficient. Reported timing values refer to different
measurement boundaries and should not be treated as directly comparable end-to-end latencies.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 17 of 40
4.1. Agricultural Workloads and Computing Platforms
Agricultural edge workloads span time-series analysis, image classification, detection,
segmentation, and quantitative spectral analysis. Low-frequency soil and environmental
data typically require preprocessing, state estimation, and periodic decisions. Crop disease
recognition, fruit detection, and navigation process higher-dimensional images; continuous
video adds sustained throughput and latency requirements [104]. Multispectral and hyper-
spectral analysis also places substantial demands on buffering and memory access. Data
dimensions, update rate, access patterns, deadlines, and power budget together define the
computing platform requirements.
Computing architectures serve different workload profiles. MCUs support lightweight
sensing and control; embedded CPUs offer flexible preprocessing and model execution.
Graphics processing units (GPUs) and neural processing units (NPUs) accelerate parallel
operations when operator support and data movement fit the software framework. FPGAs
provide configurable pipelines, local buffers, and data paths within their logic and memory
budgets. ASICs allow further optimization for stable workloads, with corresponding design
and fabrication investment [105]. Matching these strengths to the task makes latency, energy,
flexibility, and development cost explicit selection criteria.
Agricultural deployments illustrate three complementary routes to efficient edge pro-
cessing: local services, model optimization, and dedicated hardware. Adebayo et al. used an
ESP32-S3 and TinyML for environmental monitoring and control in mushroom cultivation;
detailed latency and energy characterization remains to be reported [102]. Premkumar and
Sigappi placed soil moisture prediction and irrigation control on a Raspberry Pi, reducing
average service latency from 1415.8 ms in the cloud to 313.6 ms at the edge [101]. Xu and
Liao’s optimized YOLOv11 reached 7.2 FPS for weed detection on a Jetson Nano [100]. Park
et al. implemented convolutional feature extraction and nearest-neighbor classification on
FPGAs and achieved an F1 score of 0.933 for strawberry ripeness classification [96]. The
examples locate the source of each benefit at the service, model, or circuit level and show
why training and inference platforms should be reported separately.
4.2. Model Compression and Hardware Adaptation
Quantization, pruning, and knowledge distillation reduce the memory and compu-
tational demands of edge models. Quantization lowers weight and activation precision
and exploits supported low-precision operations. Structured pruning removes channels
or layers in forms readily mapped to conventional hardware, while knowledge distillation
transfers information from a teacher to a smaller student model [106]. Figure 4c combines
relational knowledge distillation with FPGA acceleration for plant disease recognition. The
resulting latency and energy gains depend on how the compressed model maps to operators
and memory accesses on the target platform. This mapping connects model reduction to
measurable hardware efficiency.
Hardware adaptation turns a compressed model into an efficient executable. Input
resolution, conversion, operator mapping, numerical precision, and memory allocation
determine execution behavior. Unsupported operations may invoke software fallback or
data conversion, so deployment profiling should identify these costs. Figure 4a links model
development to hardware adaptation and testing. Xu and Liao’s Jetson Nano deployment
illustrates this process for optimized YOLOv11 weed detection [100]. Comparing baseline
and optimized models on the same hardware, input, and software stack isolates the benefit
of compression. Accuracy, latency, peak memory, and measured power should be reported
with the inclusion of acquisition, preprocessing, and postprocessing stated explicitly.
Task-sensitive compression preserves the features that matter for agricultural decisions.
Early disease symptoms, small pests, and occluded targets may be particularly sensitive
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 18 of 40
to pruning and low-bit arithmetic. Class-specific recall, false negatives, and tests across
independent acquisition batches or field conditions capture these effects [107]. Model
storage, version management, runtime memory, and update communication also enter the
resource budget of periodically updated nodes. A useful compression strategy therefore
meets task-specific detection requirements while reducing measured computing, memory,
and energy costs on the deployment platform.
4.3. FPGA Pipelines and Dedicated Hardware Implementations
FPGAs provide configurable logic and on-chip memory for task-specific parallel and
pipelined processing. Line buffers and sliding windows reuse pixels during filtering, seg-
mentation, and convolution. Systolic multiply–accumulate arrays improve convolution
throughput, and fixed procedures such as support vector machine (SVM) inference can
be implemented as dedicated intellectual property (IP) cores. Figure 4b,d show how
ARM–FPGA systems divide agricultural image processing among the processor, mem-
ory interfaces, and programmable logic. Hardware description languages (HDLs) and
high-level synthesis (HLS) offer implementation routes. Operating frequency, logic and
memory utilization, bandwidth, throughput, and measured power then characterize the
realized design.
FPGA studies demonstrate acceleration across neural networks, conventional classifiers,
and image preprocessing. Park et al. combined a systolic convolutional feature extractor with
dedicated k-nearest-neighbor (k-NN) logic on Cyclone IV and MAX 10 FPGAs, achieving
an F1 score of 0.933 for strawberry ripeness classification [96]. Pavani and Beulet used HLS,
block RAM (BRAM), and parallel processing to implement an SVM core for crop yield
classification. Prediction time was 96.43% shorter than CPU execution, with reported power
consumption of 1.657 W [97]. Gao et al.’s dark-channel-prior dehazing pipeline processed
1920 × 1080 video at 60 Hz on a Xilinx XC7S25 with reported power below 2 W; downstream
crop recognition was outside the reported evaluation [98]. Antora et al. implemented RGB
filtering, thresholding, and pixel counting on a Cyclone IV with external memory [99]. Each
result demonstrates a task-specific processing capability, with its input and measurement
conditions defining the basis for comparison.
FPGA design balances parallelism, numerical precision, and memory access within
available logic, digital signal processing (DSP) units, and bandwidth. Increased parallelism
raises throughput and resource demand; lower precision reduces arithmetic cost within the
task’s accuracy tolerance. Reports should distinguish synthesis and place-and-route esti-
mates from board measurements and identify whether power includes processors, memory,
and interfaces. An FPGA prototype also provides a basis for evaluating further specializa-
tion. Moving to an ASIC requires architecture refinement, technology mapping, timing and
power design, functional verification, and design for test [108]. Its expected benefits can
then be assessed against measured task performance and the additional development cost.
4.4. Memory Organization and Dataflow Optimization
Memory organization is a primary determinant of efficiency in image and spectral
processing. Repeated access to images, spectral bands, and intermediate features creates
demands for on-chip buffering and external bandwidth. When data delivery falls below the
processing rate, parallel units remain idle and throughput becomes memory-bound [109].
Coordinating registers, local buffers, and external memory with the processing architecture
makes data movement an explicit optimization target.
Dataflow design improves efficiency by retaining reusable information near the pro-
cessing units. Line buffers and sliding windows serve local image operations, while tiling
reuses features, weights, and intermediate results in on-chip memory [110]. Spectral band
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 19 of 40
selection, dimensionality reduction, and compression reduce storage and processing when
they retain the information required for classification or quantitative prediction. Buffering
and flow control align acquisition, preprocessing, and computing rates. Memory utilization,
transferred data volume, latency, and task error together quantify the benefit of these choices.
Near-sensor, in-sensor, and in-memory computing reduce data movement at different
points in the measurement chain. Near-sensor designs place preprocessing beside the
sensor; in-sensor designs use device responses to compute; and in-memory designs perform
selected operations within storage structures. Guan et al. demonstrated a programmable
superconducting nanowire single-photon detector (SNSPD) array for in-sensor prepro-
cessing, classification, and spectral analysis, achieving 92.22% accuracy in a three-class
letter-recognition task at approximately 2.3 K [103]. Although this study illustrates the
potential of photon-efficient in-sensor computing, its cryogenic operating requirement
and associated cooling infrastructure substantially limit its suitability for conventional
agricultural field deployment. It should therefore be regarded as a transferable labora-
tory demonstration rather than a field-ready agricultural sensing solution. Evaluating its
practical applicability would require system-level assessment of cooling energy, periph-
eral electronics, environmental compatibility, and total implementation cost. Figure 4e
illustrates a different route through coordinated processing clusters, on-chip memory, and
routing [111]. For agricultural adaptation, evaluating peripheral readout, conversion, mem-
ory access, and operating requirements identifies where these architectures can deliver a
net system benefit.
4.5. Matching Computational Performance to Application Requirements
Meaningful comparison of agricultural edge computing platforms requires consistent
reporting of task performance, input characteristics, resource consumption, and measure-
ment boundaries. As summarized in Table 3, input resolution or data dimensions, numerical
precision, and batch size should be specified wherever applicable. Computational perfor-
mance should distinguish inference latency, processing throughput, service response time,
and end-to-end latency, which represent different stages of operation and are not directly
interchangeable. Power consumption should be reported with its measurement method
and coverage, distinguishing chip-level estimates or measurements from board-level and
complete-system power. Likewise, model size, hardware resource utilization, memory ca-
pacity, and peak runtime memory footprint should not be treated as equivalent quantities.
The reviewed implementations also differ substantially in their level of agricultural
validation. For example, the FPGA-based field image monitoring system was evaluated
under both laboratory and outdoor conditions, whereas several other studies primarily
demonstrated processing capabilities using agricultural datasets or prototype hardware.
Edge-based irrigation control and TinyML-enabled cultivation monitoring provide examples
of agricultural service and environmental monitoring applications, although comprehensive
power, memory, and long-term reliability measurements remain limited. These distinc-
tions are particularly important when interpreting reported performance improvements.
A reduction in algorithm execution time does not necessarily reduce the total response
time when acquisition, communication, or actuation dominates the workflow. Similarly,
chip-level power figures cannot establish battery lifetime without accounting for peripheral
devices and operating schedules. The cryogenic SNSPD implementation further illustrates
why technological capability and field deployability must be assessed separately. Con-
sequently, platform selection should consider not only computational performance but
also system-level energy demand, implementation cost, environmental adaptability, and
maintenance requirements.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 20 of 40
Independent field validation connects computing performance to reliable agricultural
decisions. Illumination, cultivar, growth stage, observation distance, and installation can
shift the input distribution. Splitting data by plants, acquisition batches, locations, or
periods, as appropriate, tests these changes while preventing training–test leakage [112].
Acquisition equipment, environmental conditions, and the intended application range
provide essential context. During sustained monitoring, data quality checks, recalibration,
or model updates can address drift and changing inputs, with storage, communication, and
energy included in the update budget [113]. These measures complement timing tests by
establishing the conditions under which inference remains useful.
Agricultural computing optimization is most effective when task accuracy, latency,
energy, and resource use are evaluated together. Model compression, specialized pipelines,
and dataflow design each address a different execution bottleneck. Matching them to sensing
outputs, interfaces, power supply, and operating schedules connects component gains to
the complete workflow. This coordination across sensing, readout, storage, computing, and
communication motivates the integration strategies in Section 5.
5. System Integration of Sensing and Computing
System integration connects sensing devices, readout circuits, memory, processors,
and communication modules through electrical, thermal, and mechanical interfaces. Agri-
cultural implementations range from printed circuit boards (PCBs) and multichip modules
(MCMs) to system-in-package (SiP), MEMS and optoelectronic integration, and advanced
three-dimensional interconnects and chiplet architectures. These approaches differ in
integration density, manufacturing complexity, cost, environmental robustness, and main-
tainability. Figure 5 illustrates representative technologies, while Table 4 distinguishes
agriculturally evaluated systems from transferable implementations developed in other
engineering fields. Beyond integration density and functional coordination, their suitability
for agricultural deployment must be assessed in relation to packaging materials, operating
conditions, service life, and life-cycle cost.
Table 4 compares the quantitative characteristics and validation contexts of the reviewed
implementations. Four evidence categories organize their agricultural applicability: agricul-
tural field validation, agricultural laboratory or real-sample validation, agriculture-oriented
prototyping, and transferable non-agricultural demonstrations. The soil monitoring system
in [114] was evaluated under rice-field conditions. The wireless soil moisture sensor in [115]
demonstrates agriculture-oriented integration, with long-term field reliability remaining a
qualification priority. Studies in [82,116–119] demonstrate optical, MEMS, photonic and
three-dimensional integration capabilities that may inform agricultural hardware design.
Their classification identifies the experimental context supporting each capability and the
agricultural validation needed for its application. Measurement errors, optical characteris-
tics, circuit specifications and fabrication dimensions are interpreted within their respective
technical contexts. Agricultural measurement accuracy, complete-system power and long-
term reliability are assessed from the reported experimental evidence, with unreported
outcomes explicitly identified.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 21 of 40
Table 4. Quantitative characteristics and validation scope of representative agricultural sensing systems and transferable integration technologies.
Application and Implementation Integration Architecture Reported Quantitative Results Validation Scope and Limitations
Precision soil monitoring; ISFETs, soil
sensors, and dsPIC MCU [114]
PCB integration with UART
communication and USB-connected host
Temperature and moisture errors below 1%;
pH error 3%; NO3 − and K+ errors
approximately 17% and 14%, respectively
Prototype characterization and rice-field
measurements; spatial interpolation performed on an
external host; long-term autonomous operation NR
On-chip NIR spectroscopy; plasmonic
filters and InGaAs photodetectors [82]
Direct integration of Au/SiNx filter arrays
with photodetectors; computational
reconstruction
16 channels; 900–1700 nm; approximately 5
nm filter FWHM; computational separation of
simulated spectral peaks spaced 1 nm apart
near selected wavelengths.
Fabricated optical chip and computational
reconstruction demonstrated; direct
agricultural-sample validation and complete-system
power NR
Wireless soil moisture sensing; DPHP
sensor and mixed-signal CMOS IC [115]
MCM combining a micromachined probe
with a 0.7 μm CMOS interface and ASK RF
transmitter
12-bit digital readout; 433.92 MHz RF carrier
Agriculture-oriented integrated prototype;
quantitative soil-measurement error, field operating
lifetime, and complete-system energy NR
Monolithic CMOS–MEMS integration;
embedded MEMS switches [116]
MEMS structures fabricated within CMOS
back-end interconnect layers
Comparable agricultural sensing accuracy and
operating metrics NR
Monolithic fabrication and MEMS operation
demonstrated outside agriculture; agricultural
deployment and lifetime NR
Miniaturized FBG interrogation; III–V/Si
photonic chip [117]
BCB-bonded VCSEL, SOI waveguides,
AWG, and photodetectors
Chip footprint approximately 5 × 3 mm2 ;
wavelength resolution approximately 1 pm;
wavelength accuracy approximately ±10 pm;
temperature resolution approximately 0.1 °C
Laboratory photonic prototype; agricultural-sample
testing and field stability NR
Large-area CMOS pixel arrays;
ATLASPix3 modules [118]
Four-chip module using HV-CMOS pixel
sensors, shared interconnects, and
Shunt-LDO-based powering studies
49,104 pixels per chip; 50 × 150 μm2 pixel size;
180 nm HV-CMOS process
Non-agricultural detector prototype; multichip
readout and powering investigated; agricultural
imaging and field validation NR
3D pixel detector integration; FE-I4
readout ASIC [119]
Via-last TSVs, backside redistribution,
thinned ASIC, and flip-chip bonding
Approximately 60 μm TSV diameter; 80–100
μm final ASIC thickness
Fabricated TSV readout chips and hybrid detector
modules tested; agricultural imaging performance
and long-term package reliability NR
Notes: NR, not reported or not established for the specified evaluation. Quantitative values represent different measurement results, device specifications, or fabrication
characteristics and are not directly comparable. Agricultural field validation and long-term reliability are distinguished from laboratory device demonstrations.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 22 of 40
Figure 5. Heterogeneous integration technologies and architectures relevant to agricultural sensing
and computing systems. (a) Silicon-interposer-based 2.5D packaging, illustrating multilevel inter-
connections through microbumps, through-silicon vias (TSVs), and package substrates, together
with the fabricated package [120]; (b) wafer-level MEMS and large-scale integration (LSI), show-
ing MEMS fabrication, wafer bonding, and TSV interconnection [121]; (c) modular heterogeneous
chiplet architecture, illustrating the partitioning of monolithic systems into interconnected chiplets
through chiplet interfaces and 2.5D packaging [122]; (d) silicon photonic and microfluidic integra-
tion, featuring on-chip ring waveguide sensors, optical fiber coupling, microfluidic channels, and
electrical connections [123]; (e) multichiplet computing architecture and prototype, showing process-
ing elements, RISC-V controllers, global buffers, and interchiplet communication interfaces [124];
(f) multichannel electrochemical microfluidic sensing system, illustrating metal microelectrode fabri-
cation on glass, PDMS microchannel formation, alignment and bonding, and interfacing between
fluidic sensing structures and electronic readout modules [125]. All panels were adapted from
the cited sources under the Creative Commons Attribution 4.0 International License (CC BY 4.0;
https://creativecommons.org/licenses/by/4.0/ (accessed on 25 October 2026)). Typography and
layout were standardized without altering the original technical content.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 23 of 40
5.1. Board-Level Systems and System-in-Package Integration
Board-level integration offers a flexible platform for agricultural prototyping and
field validation. Rueda-Delgado et al. connected temperature and moisture sensors,
pH/NO3 − /K+ ISFETs, and a dsPIC33FJ128GP802 microcontroller through UART inter-
faces. USB transfer to a host computer supported spatial analysis of rice-field soil properties.
Temperature and moisture errors were below 1%, while NO3 − and K+ errors reached
approximately 17% and 14%, respectively [114]. This modular configuration supports
component selection, debugging, and expansion. Its measurement results also identify
calibration priorities for the chemical channels and define the role of external computing in
the deployed workflow.
MCM and SiP approaches bring multiple dies into a shared package to shorten connec-
tions and reduce assembly size. MCM describes multichip assembly, while SiP emphasizes
the combined system function; their implementations can overlap. Morais et al. inte-
grated a silicon-micromachined dual-probe heat-pulse (DPHP) soil moisture sensor with a
0.7 μm CMOS mixed-signal/RF chip. Signal amplification, sigma–delta conversion, and
433.92 MHz amplitude-shift keying (ASK) transmission provided 12-bit wireless read-
out [115]. The prototype demonstrates compact sensing, conversion, and communication.
Long-term package reliability and complete-system energy savings remain to be quantified;
die testing, interconnection, assembly stress, heat transfer, and manufacturing cost are the
corresponding engineering considerations.
Packaging boundaries and functional partitioning distinguish integration approaches.
A PCB connects packaged components, an MCM or SiP combines dies within a package,
and a chiplet architecture partitions functions among dies with coordinated die-to-die
communication [126]. Identifying the functions inside and outside each boundary clarifies
what the integration achieves. Manufacturing cost, replacement strategy, and long-term
reliability then determine its suitability for the agricultural application.
5.2. MEMS and Optoelectronic Integration
MEMS and readout circuits can be combined through monolithic processing or the
interconnection of separately fabricated components. Monolithic designs shorten connec-
tions and reduce parasitics when MEMS and CMOS processes are compatible. Gaddi et al.
formed microcavities and movable structures in CMOS back-end-of-line (BEOL) layers
to integrate MEMS switch arrays with electronics [116]. Separate fabrication permits in-
dependent optimization of sensing and readout, with package stress and signal integrity
handled at their interface. The Gaddi et al. study provides a manufacturing example outside
agricultural validation. These two routes offer distinct starting points for sensor integration,
to be selected according to process compatibility and measurement requirements.
Optoelectronic integration reduces discrete optical components and alignment steps
by combining filters, sources, waveguides, and detectors. Zheng et al.’s Au nanohole filters,
coupled through SiNx spacers to InGaAs photodetectors, formed a 16-channel spectrometer
spanning 900–1700 nm [82]. Li et al. used benzocyclobutene (BCB) bonding to combine
III–V vertical-cavity surface-emitting laser (VCSEL) arrays with a silicon-on-insulator (SOI)
platform. Their approximately 5 × 3 mm2 fiber Bragg grating (FBG) interrogation chip inte-
grated waveguides, arrayed waveguide gratings (AWGs), and photodetectors and achieved
a reported wavelength resolution of approximately 1 pm [117]. Detector-level integration
and heterogeneous photonic assembly provide complementary routes to compact optical
measurement. Studies in [82,117] demonstrate device-level optical capabilities that offer
a basis for agricultural adaptation. Application-specific evaluation should characterize
sample-dependent responses, optical coupling, thermal sensitivity, calibration stability and
external instrumentation requirements under representative agricultural conditions.
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 24 of 40
Microfluidic integration couples controlled sample delivery to optical or electrochemi-
cal detection. Figure 5d connects silicon photonic ring-waveguide sensors to optical fibers,
glass/PDMS channels, and PCB interfaces [123]. Gonzalez-Losada et al. developed an
eight-channel electrochemical biosensor combining metal microelectrodes on glass, plasma-
bonded PDMS channels and dedicated readout circuitry (Figure 5f) [125]. The system was
evaluated using synthetic miRNA, providing a basis for further assessment with agricultural
sample matrices and crop-pathogen targets. These examples show how fluid transport and
multichannel measurement can be coordinated. Agricultural implementations require channel
and sealing designs suited to complex matrices, bubbles, and contamination, with fluidic and
electronic regions arranged to support repeatable measurements and maintenance [127].
5.3. Three-Dimensional Interconnects and Chiplet Architectures
Through-silicon vias (TSVs), flip-chip bonding, and interposers provide compact,
high-density connections among sensors, readout, and computing dies. Figure 5a shows
a 2.5D silicon-interposer package with microbumps, TSVs, and substrate interconnects.
Chen et al. modeled its thermomechanical and electrical behavior and optimized TSV
diameter, spacing, and insulation thickness [120]. Fritzsch et al. demonstrated a via-last
process using an 80–100 μm thick readout ASIC, approximately 60 μm TSVs, backside
copper redistribution, and flip-chip connection to a pixel sensor [119]. These modeling
and fabrication examples establish transferable interconnect approaches. Agricultural
imaging and multisensor systems are candidate applications when their bandwidth and
size requirements justify the package complexity; direct agricultural validation was outside
these studies.
Chiplets partition processing, memory, and interfaces into dies linked through ad-
vanced packaging. Mbua et al. examined how workload partitioning affects communication
latency, network congestion, and throughput (Figure 5c) [122]. Shao et al.’s multichiplet
prototype combines processing elements, RISC-V controllers, global buffers, and die-to-die
interfaces (Figure 5e). These general-purpose studies connect architectural partitioning to
data exchange and hardware implementation. High-resolution imaging and hyperspectral
processing provide plausible agricultural workloads because of their computing and mem-
ory demands. For low-data-rate monitoring, simpler architectures provide the appropriate
reference against which to assess chiplet costs and benefits.
Advanced packaging is most useful when its interconnect benefits exceed the costs
of communication, assembly, testing, and maintenance [128]. Die-to-die transfers and
buffering consume time and energy, while thinning, thermal expansion mismatch, and
package stress affect reliability [129,130]. Processor and memory heat also matter when
adjacent sensing elements are temperature-sensitive. Low-frequency soil and environmental
monitoring can often use board-level or single-chip solutions. High-bandwidth imaging and
heterogeneous sensing are stronger candidates for advanced packaging when measurements
demonstrate a useful gain in size, data transfer, or coordination.
5.4. Integration Choices and Engineering Constraints
The dominant system bottleneck provides a practical rule for choosing integration.
Drift, contamination, or unstable calibration directs effort toward sensing interfaces, pro-
tection, and calibration. Interface latency, connection complexity, or data-transfer limits
motivate improved PCB connections, MCMs, or SiP . Space-constrained systems require
deliberate placement of integrated functions alongside accessible optical, fluidic, power,
and communication interfaces [131]. Advanced interconnects and chiplets become can-
didates for high-bandwidth imaging and real-time multimodal workloads. Measured
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 25 of 40
performance, manufacturing and testing cost, and maintenance requirements establish the
appropriate choice.
Selecting an integration technology for agricultural deployment requires balancing
integration density and interconnect performance with environmental resilience, cost and
maintainability. Table 5 compares representative approaches across agricultural validation,
packaging materials, environmental constraints, cost and maintenance requirements. Board-
level and modular architectures support component replacement and calibration, while
MCM/SiP and monolithic integration can reduce system size and interconnect complexity.
These more integrated approaches also place greater demands on assembly and repair.
TSV-based packaging and chiplet architectures provide additional interconnect capabilities
whose value should be assessed against workload demands and manufacturing, testing
and thermal-management requirements. Photonic and microfluidic integration introduces
application-specific requirements for optical coupling, material compatibility, fluidic sealing
and contamination control. Evaluating these trade-offs under representative agricultural
conditions guides the adaptation of laboratory demonstrations to deployment.
Table 5. Engineering suitability of representative integration technologies for agricultural sensing
systems.
Integration
Approach
Agricultural Evidence
and Potential
Applications
Packaging Materials and
Environmental
Constraints
Cost and
Maintainability
Operating Temperature
and Lifespan
Considerations
PCB-level and
modular
integration
Agricultural field
measurement
demonstrated [114];
distributed soil and
environmental
monitoring
PCB substrates,
connectors, protective
coatings; sensitive to
moisture ingress,
corrosion, and vibration
Relatively accessible
assembly and
replacement; recurring
connector and
enclosure maintenance
Verify component
temperature ratings,
sealing integrity, and
long-term connection
stability
MCM/SiP
Agriculture-oriented
wireless soil sensor
prototype [115]; compact
sensing and readout
Die attachment, package
substrates, encapsulants,
and interconnects;
thermal expansion and
moisture protection
Additional packaging
and testing overhead;
internal dies are
generally not
field-replaceable
Thermal cycling, package
integrity, calibration
retention, and service life
require validation
Monolithic
MEMS–CMOS
integration
General-purpose
fabrication demonstrated
[116]; potential for
compact sensing modules
Silicon and CMOS
interconnect layers;
exposed MEMS
structures may require
stress isolation and
environmental protection
Process compatibility
and fabrication cost are
important; device-level
replacement may be
necessary
Temperature-dependent
mechanical response,
contamination, fatigue,
and long-term drift
require evaluation
Photonic and
optofluidic
integration
Laboratory
demonstrations
[82,117,123,125];
potential for spectral and
biochemical sensing
Silicon/III–V photonics,
bonding polymers,
glass/PDMS interfaces;
optical alignment,
leakage, and
contamination
Specialized fabrication
and optical packaging;
replaceable fluidic
interfaces may reduce
maintenance burden
Thermal spectral shifts,
bonding stability, sealing
durability, and
calibration retention
require testing
2.5D/3D TSV
integration
Non-agricultural
interconnect
demonstrations
[119,121]; potential for
high-bandwidth imaging
Silicon interposers,
copper interconnects,
bonding layers;
thermomechanical stress
and package integrity
High fabrication and
testing complexity;
package-level
replacement generally
required
Thermal cycling,
interconnect fatigue, heat
dissipation, and lifetime
reliability require
qualification
Chiplet-based
integration
General-purpose
architectural and
hardware demonstrations
[122,124]; potential for
data-intensive edge
computing
Die-to-die interconnects,
advanced substrates,
redistribution layers;
thermal coupling and
interconnect reliability
Additional design,
assembly, and testing
requirements; limited
field-level repairability
Package temperature
gradients, interconnect
reliability, and operating
lifetime remain to be
evaluated for
agricultural use
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 26 of 40
Operating temperature and service life are system-level qualification criteria for each
integration approach. Evaluations should report operating temperature limits, humid-
ity exposure, thermal cycling conditions and continuous-operation duration. Reliability
assessment should distinguish electrical and package integrity from sustained measure-
ment performance, tracking baseline drift, calibration retention, interconnect integrity and
maintenance intervals. Together, these measurements establish the evidence needed for
agricultural qualification and identify the testing required to advance candidate technolo-
gies toward field validation. Final selection should balance demonstrated measurement
and computing gains against full life-cycle costs, including fabrication, assembly, protection,
calibration, replacement and maintenance.
Closer integration makes power, signal, and thermal coordination part of measure-
ment design. Digital switching and wireless transmission can couple into weak-signal
AFEs through shared supplies, grounds, and neighboring interconnects [132]. Processor,
radio-frequency, and power-management heat can shift temperature-sensitive sensing re-
sponses. Coordinated power-up, stabilization, and acquisition schedules support valid
measurements across multiple power domains [133]. Analog–digital power partitioning,
grounding and decoupling, sampling schedules, and thermal isolation or compensation
address these couplings. Measurements of noise, baseline, thermal response, and stability
under representative computing and communication loads establish their effectiveness.
Modular integration can accommodate the different service lives of sensing interfaces
and electronics. Electrodes, membranes, and microfluidic components exposed to samples
may need cleaning, recalibration, or replacement, while readout and computing modules
can be reused [134]. Detachable electrical and fluidic interfaces should therefore be designed
for repeatable connection, sealing, and calibration within the available size budget. Edge
hardware that receives model updates also needs sufficient configurability and software
compatibility. This approach links integration to sustained measurement performance and
life-cycle cost.
6. Stability and Maintenance for Field Applications
Sustained agricultural sensing depends on retaining measurement quality through
environmental exposure and maintenance. Temperature, humidity, soil and liquid contact,
particle deposition, and mechanical disturbance can affect circuits, package interfaces, opti-
cal paths, and calibration. Reliability evaluation therefore covers both hardware operation
and the stability of the information it produces [134]. This section connects environmental
effects to drift control, protective packaging, maintainability, and field validation. The
deployment conditions and observation period define the reliability evidence relevant to
each application.
6.1. Environmental Effects and Measurement Drift
Environmental records and reference measurements help separate changes in agri-
cultural targets from changes in the instrument. Temperature affects sensing materials,
analog circuits, and photodetectors; humidity and condensation affect insulation. Dust can
attenuate optical signals or obstruct channels, while soil solutions and complex matrices
can foul membranes and destabilize references [135,136]. Recording relevant conditions
alongside blanks, reference responses, or periodic calibrations makes these effects traceable
and supports interpretation of field measurements.
Stability can be improved at the sensing interface and within the readout architecture.
Grygołowicz-Pawlak et al. used a functionalized monolayer with ferrocene and terminal
acrylate groups to covalently connect sensing layers in solid-state ion-selective sensors, re-
ducing potential drift to 0.2 mV/day while retaining selectivity and calibration slopes [137].
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 27 of 40
Jones et al.’s portable fiber-optic pH sensor combined dual-wavelength LEDs, silicon pho-
todiodes, reference channels, and internal calibration to reduce light-source fluctuation
effects [138]. The latter demonstrates optical compensation in the tested configuration, with
long-term field stability still to be established. For plant-mounted sensors, growth, defor-
mation, and changing contact add interface requirements. Baseline, sensitivity, reference
response, and contact integrity provide complementary measures of stability.
Environmental testing links specific stresses to changes in device performance. Thermal
cycling produces stress through expansion mismatch; condensation, vibration, and repeated
contact can alter insulation, connections, and optical coupling [139]. Materials, packag-
ing, and installation determine whether the response is reversible or reflects permanent
degradation [140]. Individual and combined stress tests should therefore compare baseline,
sensitivity, noise, references, and connection integrity before and after exposure. Continu-
ous field operation adds evidence of performance under interacting stresses; service-life
estimates require a validated relationship between those conditions and degradation.
6.2. Protective Packaging and Maintainable Design
Protective packaging preserves the measurement pathway while shielding electronics
from moisture, contamination, and mechanical damage [141]. Electrochemical devices
need exposed sensing surfaces with reliable insulation; optical devices need clean, stable
windows; microfluidic devices need sealed channels and electrical isolation. Buried probes
add prolonged moisture contact and mechanical loading [142]. Coatings and encapsulants
can also influence measurements through adsorption, permeability, and interfacial effects.
Comparing sensitivity, baseline, repeatability, and sealing before and after packaging verifies
that protection supports the intended measurement.
Stable attachment supports repeatable plant-wearable measurements. Wei et al. devel-
oped a plant-compatible TIP-adhesive based on 2-hydroxy-3-phenoxypropyl acrylate (HP A),
with adhesion in approximately 2 s and strong attachment across tested temperatures of
5–45 °C and relative humidities of 35–95%. Leaf morphology and stomatal characteristics
were used to assess compatibility, and wearable sensors monitored leaf elongation con-
tinuously and noninvasively [143]. The work establishes attachment performance within
these tested conditions. Extending that capability to field systems requires connections
that accommodate leaf movement and growth, together with verification of measurement
repeatability throughout use.
Maintainable designs make cleaning, replacement, and recalibration accessible through-
out deployment. Electrodes, membranes, and channels may need periodic service, while
optical devices require access to windows and calibration procedures [144]. Replaceable
sensing and fluidic components can reduce recurring cost when reassembly preserves
electrical contact, sealing, and calibration. Baseline, sensitivity, reference responses, and
repeatability before and after intervention quantify recovery [145]. Recording service
intervals then connects maintenance effort to sustained measurement quality.
6.3. Field Validation and Performance Retention
Reliability evidence is most informative when the tested hardware, conditions, operat-
ing mode, and duration are explicit. Bare-device tests characterize sensing structures and
electrical behavior. Packaged-module tests add protection and interconnection. Complete-
node tests cover sensing, computing, communication, and power management. Controlled
laboratory operation identifies drift, stress tests probe degradation mechanisms, and field
deployments capture environmental variation, installation, and maintenance together [146].
Combining these levels supports increasingly specific operational claims. Long-term service-
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 28 of 40
life estimates additionally require adequate duration, sample coverage, and validated life-
time models.
Long-term field records reveal how measurement accuracy, data continuity, and main-
tenance interact. Gasch et al. monitored 210 soil moisture sensors across 42 locations and
five depths over approximately nine years, reporting an average missing-record propor-
tion of 35% across sensors [147]. The network required 68 sensor replacements, 33 cable
repairs, and five data logger replacements, illustrating the operational consequences of
sensor degradation, physical damage, and power interruptions. These findings support
the use of protected interconnections, accessible power and logging modules, and targeted
redundancy at critical monitoring locations. Redundant or reference sensors should be
distributed across representative soil conditions and depths, while independent logging
or power paths may be required where common-mode failures threaten data continuity.
Maintenance schedules should account for battery life, historical failures, environmental
exposure, and agricultural operations, with additional inspections triggered by persistent
data loss or abnormal sensor behavior.
Data quality management should distinguish missing observations, invalid measure-
ments, and estimated values. For each monitoring interval, data completeness can be
defined as the number of quality-accepted observations divided by the number of sched-
uled observations during active deployment. The longest consecutive data gap, frequency
of interruptions, and proportion of flagged measurements provide complementary indica-
tors of network performance. Range violations, persistent flat-line responses, spikes, and
abrupt signal changes can be used to identify records requiring inspection or exclusion.
Short gaps may be estimated using validated temporal or spatial relationships when the
target dynamics permit, whereas extended gaps or periods involving rapid rainfall and
irrigation changes should not be filled without sufficient supporting evidence. Original
measurements, quality-control flags, and estimated values should be retained separately,
with imputation methods and associated uncertainty documented.
Recalibration should be triggered by persistent deviations from independent reference
measurements, changes in sensor sensitivity, or systematic drift exceeding application-
defined tolerances, rather than by elapsed time alone. Before recalibration, power supply,
cable integrity, sensor contact, and environmental interference should be examined to dis-
tinguish instrument faults from genuine changes in the measured agricultural variable.
Reference measurements and repeatability checks before and after maintenance can ver-
ify recovery of measurement quality. Environmental qualification should also address
moisture ingress, thermal cycling, mechanical disturbance, and degradation of exposed
sensing interfaces [148,149]. Field evaluations should report measurement error, drift, data
completeness, maximum outage duration, maintenance and recalibration frequency, and
performance recovery after intervention. Together, these indicators provide a traceable
basis for assessing long-term measurement reliability and the practical value of protective
and modular hardware designs.
7. Research Priorities for Agricultural Sensing and Computing
The technologies reviewed here provide complementary building blocks for task-
specific agricultural hardware. The next design opportunity is to coordinate their con-
tributions across acquisition, readout, computation, and sustained operation. Figure 6
organizes optical, electrochemical, and particle-sensing interfaces with computing, mem-
ory, communication, and power management. It supports three priorities: joint design of
sensing and computation, reuse of compatible modules at an appropriate integration level,
and traceable calibration with comparable evaluation. The framework guides functional
https://doi.org/10.3390/agriculture16222497

Agriculture 2026, 16, 2497 29 of 40
partitioning across devices, packages, and boards; each proposed configuration can then be
tested against its agricultural task requirements.
Figure 6. Sensing–computing integration framework, environmental constraints, and future directions
for smart agriculture.
7.1. Task-Driven Sensing–Computing Co-Design
Joint sensing–computing design begins with the information needed for the agricul-
tural task. Spectral bands, spatial resolution, regions of interest, and sampling frequency
determine acquisition volume, readout bandwidth, memory, and processing. For disease
detection and hyperspectral analysis, selecting informative bands or spatial samples at
acquisition can reduce both data collection and subsequent transfer; selection after full
acquisition instead reduces downstream processing [150]. The design objective is to retain
features needed for early symptoms, low-abundance targets, and transient events while
reducing redundant work. Tests across representative samples and conditions should
quantify this balance through task performance, end-to-end latency, and energy.
Table 6 translates these principles into representative hardware configurations for three
agricultural tasks with distinct system bottlenecks. Soil nutrient monitoring prioritizes
selective sensing, stable analog readout and long-term calibration. Plant phenotyping
requires coordinated data acquisition, memory managemen

论文原文与出版信息

Agriculture
查看论文原文

继续探索