A TinyMLOps Pipeline for Coarse-Grained Plant Disease Classification in Precision Agriculture

Logical Operator Operator

Search Text

Search Type

add_circle_outline

remove_circle_outline

Article Menu

Font Type:

Arial Georgia Verdana

Font Size:

Aa Aa Aa

Line Spacing:

  

Column Width:

  

Background:

Open Access Article

by
Hossein Aqasizade

1,2,
Mattia Antonini

1,
Massimo Vecchio

1,* and
Fabio Antonelli

1

1

Fondazione Bruno Kessler, 38123 Trento, Italy

2

Department of Information Engineering and Computer Science, University of Trento, 38123 Trento, Italy

Author to whom correspondence should be addressed.

Submission received: 23 May 2026 / Revised: 31 July 2026 / Accepted: 4 August 2026 / Published: 7 August 2026

Abstract

Identifying plant health conditions is an emerging precision-agriculture and food-security challenge, intensified by deploying deep-learning models on memory- and power-constrained edge devices. We present a TinyMLOps pipeline spanning model design, optimization, quantization, and deployment across diverse edge devices, evaluated under controlled laboratory conditions. Using a dataset derived from the PlantVillage benchmark, 39 fine-grained classes are aggregated into three superclasses: healthy leaf, unhealthy leaf, and no leaves. The resulting system therefore performs plant health-status classification and background filtering rather than diagnosing specific diseases. We train a MobileNet-based convolutional neural network jointly optimized for classification accuracy and computational efficiency, adopting state-of-the-art hyperparameter optimization (HPO) tools. Five models are selected, four from the Pareto Front and one as the biggest evaluated model during HPO, converted to LiteRT and ONNX, and evaluated at

float32
and post-training

int8
precision on a Raspberry Pi Zero 2 W and an STM32H743ZI microcontroller. At

float32
, LiteRT is 1.87–2.65× faster than ONNX Runtime on the Raspberry Pi across all five models. Relative to their

float32
LiteRT counterparts, the

int8
LiteRT models are 2.83–3.67× smaller on disk and 21.3–31.2% faster on the same board, at a cost in F1-score of between 0.0010 and 0.0068. On the microcontroller, only the two smallest models deploy at both precisions; for these, the fully quantized

int8
variants are

$4.3 \times$
faster and 3.85× smaller in MCU flash footprint than the

float32
counterparts. The mid-range model fits the 2 MB flash and 1 MB RAM budget only when quantized, while the two largest models exceed it in every configuration tested.

1. Introduction

As the need for food security grows, precision farming stands as a powerful ally, leveraging the most cutting-edge technologies to enhance crop production. Every year, significant portions of farm production are wasted due to recurring pests, devastating diseases, harsh weather patterns, and subtle soil erosion, all of which contribute to these losses. The precision farming paradigm, enabled by modern technology to make farming more efficient and productive, provides a powerful answer to these growing threats. To protect plants and secure the farming economy, the ability to make early, accurate plant disease identification remains fundamental, as it enables quick action before a widespread epidemic. Traditionally, this function was provided by biological methods, as in the case of “sentinel” plants [1], which exhibit symptoms indicative of particular pathogens, serving as a sort of natural early warning system. Although theoretically sound, traditional methods are bogged down by practical constraints. They are quite time- and labor-intensive to plant, grow, and continuously monitor. Further, human visual inspection is prone to error, particularly in misinterpreting or completely overlooking disease symptoms [2]. As a countermeasure to the limitations of manual, biology-focused monitoring, the farming industry continues to embrace technological innovations. Digital farming leverages solutions such as computer vision, Internet of Things (IoT) sensor systems, and AI-driven predictive analytics to provide faster, scalable, and more reliable real-time monitoring of crop health [3]. Among the array of methods, the combination of computer vision and deep learning has proven one of the most effective for mechanizing plant disease classification from images [4].

The noteworthy aspect in this case is the direct execution on edge devices, from microcontrollers (MCUs) and single-board computers (SBCs) to real-time, in-situ processing of a continuous stream. Such embedded settings are themselves defined by tight computational constraints that profoundly impact model design, from architectural size (megabytes vs. kilobytes) to data types (

float32
vs.

int8
), among others. In real agricultural deployments, connectivity is often intermittent or entirely unavailable, and power budgets are limited. For this reason, disease detection is better performed locally, directly on embedded devices, enabling real-time responses while minimizing communication, energy consumption, and operational cost. Systematically managing these restrictions is necessary, and the TinyMLOps approach provides a systematic methodology for model design, optimization, and execution on devices with limited capabilities [5].

This work describes the development of a computer vision system based on TinyMLOps principles, designed to run an optimized deep neural network model on embedded platforms for plant disease monitoring. The system uses the state-of-the-art MobileNetV1 backbone architecture [6] as a pre-trained feature extractor, while a custom final classification layer is trained from scratch. Optuna [7] is utilized for hyperparameter tuning, balancing detection accuracy and model complexity.

The experiments were conducted on the PlantVillage dataset, a widely used benchmark in the plant disease detection literature, consisting of high-quality leaf images acquired under controlled laboratory conditions. While this dataset supports rigorous and reproducible evaluation of the vision-sensing pipeline, it does not reflect the full complexity of outdoor deployments, including variable lighting, occlusions, and background clutter. The quantitative results reported here should therefore be interpreted as benchmark performance. To assess operability beyond PlantVillage, the system is additionally evaluated qualitatively on previously unseen, locally collected leaves (Section 4.2.3). This work contributes the systematic integration of model optimization, deployment, and hardware evaluation into a reproducible TinyMLOps lifecycle for heterogeneous edge devices in precision agriculture. The ultimate operational objective is real-time, on-device plant-health screening under field conditions. The envisioned inference pipeline for field deployment is illustrated in Figure 1.

Figure 1. Envisioned on-device inference pipeline for real-time plant-health screening under field conditions. A captured image is preprocessed and passed to the quantized model, which classifies the input as a healthy leaf, an unhealthy leaf, or no leaf.

The main contributions of this article are summarized as follows:

Designed and implemented an end-to-end TinyMLOps pipeline for edge deployment, integrating state-of-the-art lightweight CNN architectures (e.g., MobileNet) and covering the full workflow from model development to on-device inference;

Comprehensive hardware-level benchmarking and deployment validation on heterogeneous edge devices, namely a Raspberry Pi Zero 2 W and an STM32H743ZI microcontroller, providing detailed analysis of inference latency, memory footprint, and accuracy trade-offs, and establishing the feasibility of on-device inference.

Optimization and adaptation of plant disease classification models for multiple resource-constrained devices.

The remaining parts of the paper are organized as follows: Section 2 provides an overview of research in Edge AI and TinyMLOps, reviewing prior frameworks, deployments, and optimization approaches. Section 3 details the proposed workflow, including dataset preparation, model design/training, optimization, and edge-device deployment and configuration. Section 4 reports the experimental results on both target platforms, together with a discussion of the observed trade-offs, limitations, and a roadmap towards field deployment. Finally, Section 5 concludes the paper by summarizing the main findings and outlining directions for future work.

2. Related Work

Research related to this study spans multiple intersecting areas, including TinyMLOps frameworks, edge AI architectures, model and optimizations for TinyML, and plant disease detection on edge devices. This section reviews the most relevant contributions across these dimensions and positions the proposed work within the existing literature.

2.1. Vision-Based Plant Disease Detection in Precision Agriculture

Wadhwa et al. [8] propose a generalized hybrid CNN ensemble for multi-crop disease classification that uses F1-score driven Dirichlet weighting and CatBoost stacking, with SHAP-based explainability, while Sakthipriya et al. [9] propose a precision agriculture approach for rice nutrient management that uses genetic algorithms to automatically design and optimize CNN architectures. Their GENCNN framework focuses on model automation and accuracy, aiming to improve nitrogen management and reduce fertilizer costs, but does not address deployment, edge constraints, or operational lifecycle aspects. Sivakumar et al. [10] propose a computer vision–based crop and weed segmentation framework using federated ensemble learning across heterogeneous multi-crop field datasets. Their approach enables scalable, privacy-preserving model training and incremental deployment without centralized data sharing. Khan et al. [11] propose a MobileNetV3-Small-based model for plant disease classification on edge devices. Their approach focuses on achieving high accuracy on the PlantVillage dataset and applies post-training quantization to reduce model size. Several additional studies provide relevant accuracy–complexity references on PlantVillage, including classifiers based on AlexNet and GoogLeNet [12], InceptionV3 [13], a compact nine-layer CNN [14], and a modified InceptionV3 [15]. Their reported predictive metrics and parameter counts are interpreted in Section 4.4, with explicit consideration of differences in class granularity and experimental protocol. Arratia et al. [16] propose the BODOQUE (Bimodal Observational Device for Optimizing Quantification of Ephemeral streams), a systematic framework for deploying TinyML in real-time environmental monitoring that combines low-power sensing with event-triggered, high-performance processing to minimize energy consumption. Other agricultural applications emphasize model design rather than operational deployment. In [17], the authors introduced a general inference pipeline for plant disease detection running on a small single-board computer (i.e., a Raspberry Pi Zero 2 W). Building on that foundation, this paper focuses on the specific optimizations required for efficient on-device execution at the network edge. Although these studies demonstrate the effectiveness of deep learning for agricultural disease monitoring, many prioritize classification accuracy over deployment feasibility, runtime efficiency, and operational lifecycle management on embedded sensing platforms.

2.2. Lightweight Deep Learning Models for Edge Agricultural Monitoring

To go beyond hardware limits, scientists have attempted architectural and algorithmic advancements. Larian et al. [18] introduced the InTec framework and experimented with a three-layer architecture. They found that dividing the ML pipeline across the thing, edge, and cloud layers might significantly reduce latency and network traffic. In a more comprehensive study, Jenhani et al. [19] conducted an experimental investigation of split learning. On a customized testbed composed of ESP32 boards, they experimented with the efficiency of a partitioned MobileNetV2 model across different low-power wireless protocols. They identified optimal split points to balance on-device computation and communication overhead. Rather than disseminating the model, Hayajneh et al. [20] examined the potential to enhance the model’s adaptability. Their study on on-device transfer learning demonstrated its feasibility for enabling models to acquire knowledge from novel data directly at the edge. Cerioli et al. [21] expressly stated the advantage of strongly specialized optimization in their research on the detection of microplastics. They built a tailored compiler to generate native C code for their model without relying on standard inference engines. From their experimental data, they found that their technique achieved significant increases in latency, memory use, and energy efficiency compared with typical techniques. Similarly, Lê et al. [22] are interested in the run-time dynamics of TinyML and introduce a new TinyMLOps approach to real-time frame-based event classification on ultra-low-power microcontrollers. They presented a comparison using available frameworks, such as LiteRT and NNoM, to support their strategy. They even use TinyML to secure consumer IoT devices.

2.3. TinyMLOps and Edge AI Frameworks for Resource-Constrained Deployment

The Edge AI paradigm is a revolutionary shift that is driving new approaches to deploying applications on resource-limited devices, especially in verticals such as precision agriculture. The peculiar restrictions on these systems have led to the definition of Tiny Machine Learning Operations (TinyMLOps) [5]. Antonini et al. research [5] presents a comprehensive, end-to-end industrial IoT application for operationalizing anomaly detection in submersible water pumps. Additionally, this paper utilizes readily available, inexpensive hardware and presents a rigorous empirical study to support its findings. Moreover, they provide a formal, detailed description of the conceptual TinyMLOps framework, which applies classical machine learning orchestration principles to the unique domain of non-Linux-based, far-edge devices. However, Leroux et al. [23] say wide deployment is impeded by operational obstacles that transcend computation and reach to model observation, management, security, and integrity verification on distributed device populations. To address these operational concerns, various MLOps frameworks have been developed.

For an open-source alternative, Peltonen et al. suggested LinkEdge [24], an end-to-end toolset for connecting edge devices and supporting MLOps. They tested the system by creating a model to predict air-pressure failures in trucks and concluded that their open-source alternative performed as well as commercial tools while providing freedom from vendor lock-in. Extending on-device execution boundaries, Min et al. [25] created a multi-tenant sensory edge device runtime called SensiX++. They discovered that a resource-aware model server and an adaptive scheduler could be adopted to securely and efficiently run multiple isolated models on a single device. To overcome the practical hurdles across the entire workflow, Huang et al. [26] proposed the RIOT-ML toolkit. They propose a “batteries-included” package that includes a low-power OS, a model transpiler, and a secure over-the-air (OTA) update system and demonstrate its usefulness for benchmarking model performance on diverse IoT hardware. In other contributions, Karras et al. [27] contend that effective data management is a condition for successful edge AI. Towards that end, they designed and delivered a suite of TinyML-specific data-cleaning, clustering, and compression algorithms.

In other work, Katib et al. [28] designed a Deep Learning technique to secure IoT consumer devices with TinyML Driven Real-time Anomaly Detection for Predictive Maintenance (DLTML-RTADPM), to achieve real-time anomaly detection and predictive maintenance. Their approach comprises Z-score normalization as the data preprocessing method, the Fennec Fox Optimization Algorithm to select features, a bidirectional long short-term memory (BiLSTM) network for anomaly detection, and the Jaya optimization algorithm to optimize hyperparameters. Flores et al. [29] suggest a system of tiny machine learning tasks under the Internet of Intelligent Vehicles. Two ESP32 units are used in their solution, with one serving as a radio station and the other as an onboard diagnostic scanner. These updated models are then sent over a wireless channel. This system shows a marked improvement in model efficacy pos