论文信息

Akniyet Nurzhaubayev / Seilov Shakhmaran / Bibinur Zhursinbek 等 6 人

Decision Making Applications in Management and Engineering

论文解读

AI 辅助整理
方法
量化YOLOv8n边缘端+YOLOv8l云端,置信度驱动分流,自建叶片果实数据集
发现
多数图像本地处理,模糊样本上云,兼顾精度、带宽与延迟。
机会点
可探索置信度阈值自适应学习及多作物、多病害迁移的边云调度策略。

论文摘要

及时识别病害是有效果园管理的基本要求。本研究提出了一种面向苹果病害自动识别的自适应决策支持系统,该系统将深度学习与边缘-云混合计算架构相结合。所提出的框架根据预测置信度动态确定推理的位置和执行策略。为便于模型开发与评估,本研究建立了一个包含叶片和果实图像的领域专用数据集,并由专业生物学家提供标注。在配备神经处理单元(NPU)的低成本边缘设备上部署了量化YOLOv8n模型,而在云端部署了更大的YOLOv8l模型以执行进一步分析。该框架引入了一种置信度驱动的决策机制,将具有高置信度的预测保留在边缘端,将模糊案例转发至云端进行进一步评估,并丢弃置信度低于可接受阈值的观测结果。通过该机制,预测置信度被转化为可操作的决策,以指导计算资源的分配。使用独立果园测试数据集进行的实验验证表明了所提出的病害检测框架的可靠性,以及其相对于仅基于边缘或云端推理的传统方法的优越性能。该自适应架构在提高检测精度的同时,降低了带宽需求和端到端响应延迟。特别是,大多数图像在本地进行分析,而基于云端的处理则保留给预测不确定的案例。这些发现表明,置信度引导的自适应决策能够同时提升检测精度、通信效率和响应时间。因此,所提出的边缘-云决策支持方法为精准农业和物联网(IoT)应用提供了一种智能且计算高效的框架。

全文

Decision Making Applications in Management and Engineering

Decision Making: Applications in Management and Engineering, V olume 9, Issue 1 (2026) 685-713
685
Adaptive Edge Cloud Decision Support for Apple Disease Detection
Using Deep Learning
Akniyet Nurzhaubayev1, Seilov Shakhmaran1, Bibinur Zhursinbek1, Yerden Zhursinbek1, Stefan
Gubo2, Amir Mosavi2,3,4,5,*
1 Institute of Digital Sciences and Artificial Intelligence, Eurasian National University, Astana, Kazakhstan.
2 Faculty of Economics and Informatics, Selye Janos University, Komarno, Slovakia.
3 John von Neumann Faculty of Informatics, Obuda University, Budapest, Hungary.
4 Astana IT University, Astana, Kazakhstan.
5 Ludovika University of Public Service, Budapest, Hungary.
ARTICLE INFO ABSTRACT
Article history:
Received 15 November 2025
Received in revised form 19 April 2026
Accepted 17 May 2026
Available online 30 June 2026
Keywords:
Adaptive Decision-Making, Edge-
Cloud Decision Support, Orchard
Disease Detection, Decision-Making,
IoT , Edge Computing, Soft
Computing, YOLOv8, Deep Learning,
Object Detection, Artificial
Intelligence
The timely identification of diseases is a fundamental requirement for effective
orchard management. This study introduces an adaptive decision support
system for automated apple disease identification that integrates deep learning
with a hybrid edge -cloud computing architecture. The proposed framework
dynamically determines the location and execution strategy of inference based
on prediction confidence. To facilitate model development and assessment, a
domain-specific dataset comprising leaf and fruit images was established, with
annotations provided by professional b iologists. A quantised YOLOv8n model
was implemented on a cost -effective edge device equipped with a neural
processing unit (NPU), whereas a larger YOLOv8l model was deployed in the
cloud to undertake further analysis. The framework incorporates a confidence-
driven decision mechanism that retains predictions exhibiting high confidence
at the edge, forwards ambiguous cases to the cloud for further assessment, and
discards observations whose confidence levels fall below the acceptable
threshold. Through this mechanism, prediction confidence is translated into
actionable decisions governing computational resource allocation.
Experimental validation using an independent orchard test dataset
demonstrates the reliability of the proposed disease detection framework and
its superior performance relative to conventional approaches based exclusively
on edge or cloud inference. The adaptive architecture enhances detection
accuracy whilst simultaneously decreasing bandwidth requirements and end -
to-end response latency. I n particular, the majority of images are analysed
locally, with cloud -based processing reserved for cases characterised by
uncertain predictions. These findings establish that confidence-guided adaptive
decision-making can concurrently enhance detection accuracy, communication
efficiency and response time. Accordingly, the proposed Edge -Cloud Decision
Support approach offers an intelligent and computationally efficient framework
for precision agriculture and Internet of Things (IoT) applications.

Decision Making: Applications in
Management and Engineering

Journal homepage: www.dmame-journal.org
ISSN: 2560-6018, eISSN: 2620-0104

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

686

  1. INTRODUCTION
    Fruit production in orchard systems represents an important component of global agricultural
    activity, as crop health directly influences yield stability, food security and economic sustainability.
    Accordingly, orchard management involves a co mplex decision -making process characterised by
    multiple interconnected considerations [1]. Adaptive decision -making enables management
    strategies to respond promptly to variations in disease occurrence and computational demands.
    Within this context, intelligent decision systems and edge -cloud-based decision -support
    architectures offer rel iable and efficient solutions for addressing the challenges associated with
    orchard disease management [2].
    The occurrence of disease outbreaks in apple orchards can lead to substantial agricultural and
    economic losses, making early identification particularly important for enabling timely intervention
    and preventive disease control. Accurate and timely detection can also help limit the excessive
    application of chemicals and reduce the operational costs associated with orchard management [3].
    Furthermore, computer vision technologies facilitate the continuous observation and assessment of
    crop conditions in natural field environments. Their application consequently strengthens precision-
    oriented orchard management and contributes to more sustainable agricultural production practices
    [4]. Advances in deep learning and computer vision have significantly enhanced the capabilities of
    automated plant disease identification systems. In particular, YOLOv8 combines rapid inference with
    precise disease localisation, whereas lightweight convolutional neural network (CNN) architectures
    have demonstrated competitive detection performance when evaluated using curated datasets [3;
    5]. However, datasets developed under controlled experimental conditions do not necessarily reflect
    the diverse environmental characteristics encountered in operational orchards. Variations in lighting
    conditions, compl ex visual backgrounds and differences in disease appearance may therefore
    compromise the reliability of automated detection. Edge -cloud computing provides an alternative
    computational architecture through the integration of local processing capabilities and remote cloud
    resources. Promising results have been documented for edge -based agricultural detection
    applications [6]. Nevertheless, the implementation of these technologies in operational orchard
    environments continues to present considerable challenges owing to practical field constraints [7].
    This study primarily contributes an adaptive decision-making framework specifically designed for
    orchard disease detection. Through the proposed Edge -Cloud Decision Support mechanism, the
    framework dynamically identifies the most suitable computational environment for executing
    inference based on the confidence of individual predictions. Predictions that achieve an adequate
    confidence level are processed lo cally at the edge, whereas cases requiring further evaluation are
    directed to cloud -based inference. Consequently, the proposed framework broadens the scope of
    conventional disease detection by incorporating computational decision-making alongside detection
    accuracy. Rather than relying on a fixed allocation of computational resources, the system
    dynamically determines the appropriate processing strategy according to prediction uncertainty and
    prevailing operational conditions. This integration establishes an adaptive decision-making approach
    intended to facilitate the practical implementation of intelligent agricultural decision -support
    systems.
    The deployment of lightweight deep learning architectures often entails a decline in detection
    accuracy following quantisation and model compression. Consequently, implementing such models
    on edge devices may compromise the reliability of Orchard Disease Detection. Although cloud-based
    inference can offer improved detection performance through access to more powerful
    computational resources, its operation depends on network availability. Moreover, communication
    delays associated with data transmission may adversely affect real -time Orchard Disease Detection.
    These constraints are particularly pronounced in geographica lly isolated agricultural regions, where

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

687
network connectivity is frequently unreliable. In addition to computational and communication
limitations, environmental variations introduce further challenges to detection reliability. Changes in
illumination, ove rlapping leaves and interference from complex backgrounds increase visual
uncertainty, while the identification of subtle symptoms during the early stages of infection makes
Orchard Disease Detection considerably more challenging [8]. Consequently, achieving consistently
reliable disease identification under practical field conditions remains difficult.
An important research gap concerns the development of Adaptive Decision-Making strategies for
hybrid inference architectures. Most existing investigations concentrate primarily on improving
model accuracy or enhancing the computational efficiency of edge-based implementations, whereas
comparatively limited consideration has been given to Edge -Cloud Decision Support mechanisms.
Such mechanisms are required to dynamically coordinate inference execution according to prediction
uncertainty while accounting for the computational limitations of individual devices. Furthermore,
the adoption of domain-specific datasets annotated by experts remains inconsistent across existing
investigations [9]. This shortcoming may restrict the generalisation capability of trained models
across different orchard environments and climatic conditions. Accordingly, effective Adaptive
Decision-Making requires not onl y reliable disease identification but also validation within the
relevant environmental and operational context. The development of scalable hybrid architectures
capable of explicitly incorporating prediction uncertainty therefore remains an insufficiently
investigated research direction.
Recent investigations into edge computing demonstrate that deploying disease detection models
on platforms such as Raspberry Pi and Jetson Nano can facilitate real-time agricultural monitoring at
relatively low cost while maintaining acceptable detection accuracy [5]. Similarly, hybrid edge-cloud
architectures, which perform preliminary analysis locally and delegate computationally demanding
assessments to remote cloud infrastructure, have demonstrated improvements in processing
efficiency and system scalability. Further studies have identified reductions in latency, more efficient
bandwidth utilisation and improved operational resilience under unstable network connectivity in
agricultural environments [6]. Collectively, these findings establish the techn ical feasibility and
practical relevance of integrating edge and cloud resources within agricultural disease detection
pipelines. A further unresolved issue involves the development of intelligent offloading mechanisms
capable of determining whether locall y generated predictions should be accepted, transferred to
cloud-based inference or rejected. Existing approaches frequently rely on predefined static
confidence thresholds, which restrict their adaptability and may increase false -positive detections
under changing field conditions. In addition, optimisation methods, including pruning, quantisation
and knowledge distillation, are commonly investigated as separate techniques rather than being
incorporated into jointly coordinated hybrid inference architectur es [10]. Consequently, the
simultaneous optimisation of detection accuracy, processing latency and bandwidth consumption
has not received sufficient attention, despite its importance for reliable dep loyment in operational
orchard environments.
Empirical findings further suggest that combining advanced model optimisation with hybrid
inference architectures can improve the performance of agricultural monitoring systems. Fine-tuned
cloud-edge detection f rameworks have demonstrated improvements in detection accuracy and
operational efficiency across smart farming applications. Furthermore, systematic reviews highlight
the importance of adaptive, multi -stage detection strategies that explicitly account for
environmental variability and prediction uncertainty. Integration with Internet of Things (IoT)
infrastructures additionally facilitates distributed data collection and supports scalable
implementation across multiple orchard locations. Collectively, these developments provide a
rationale for designing computational architectures that integrate lightweight embedded detection

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

688
with high-capacity verification mechanisms.
Within the specific operational setting considered in this study, observations conducted i n the
Issyk Aport apple orchard are affected by complex visual backgrounds, seasonal environmental
variations and intermittent network connectivity. These characteristics necessitate the development
of domain-specific datasets and robust deployment strategies capable of accommodating local field
conditions [11]. Accurate labelling by agricultural specialists is particularly important, given that
disease symptoms may be subtle and difficult to distinguish during their initial development.
Accordingly, the study objectives focus on designing a real -time disease detection approach that
maintains high detection accuracy without compromising low -latency processing under practical
orchard conditions. The principal aim is to develop a hybrid edge -cloud computing architecture in
which a quantised YOLOv8n model operates on embedded hardware and selectively t ransfers
uncertain predictions to a high -capacity cloud-based detector according to prediction confidence.
This configuration is intended to facilitate accurate, bandwidth-efficient and real-time apple disease
detection. The timely identification of apple diseases is fundamental to minimising crop losses and
strengthening orchard management practices. Edge computing and hybrid edge-cloud architectures
are increasingly being investigated as viable approaches for achieving real -time agricultural disease
detection. In particular, integrating the Internet of Things (IoT) with artificial intelligence (AI) at the
edge enables remote crop monitoring, reduces processing latency and facilitates immediate decision-
making [1]. The application of mobile edge AI for disease identification and mapping in citrus orchards
further illustrates the advantages of performing inference locally using low -power computational
devices. Similarly, lightweight neural network architectures implemented on Arduino -based edge
platforms have demonstrated the feasibility of efficient, real -time apple disease classification at the
orchard level. Reviews examining lightweight deep learning architectures also emphasise the
importance of computationally efficient models for enabling real -time disease identification in
resource-constrained environments [3]. A broad range of deep learning architectures, including You
Only Look Once (YOLO), Faster Region -based Convolutional Neural Network ( Faster R -CNN),
MobileNet, EfficientNet and NASNet-Mobile, have been employed for fruit disease detection. Their
detection performance varies according to the selected dataset and the preprocessing procedures
applied. Architectures such as YOLOv4, YOLOv8, Inception-ResNet-V2, VGG-INCEP and AlexNet have
demonstrated different levels of detection accuracy across agricultural imaging applications. Among
these approaches, the lightweight YOLOv8n architecture offers low -latency inference on edge
devices while retaining high detection accuracy, making it a suitable candidate for deployment within
hybrid edge-cloud computing frameworks [5].
The comparative findings summarised in Table 1 indicate that previous investigations have
predominantly focused on maximising cloud-based detection accuracy or improving computational
efficiency at the edge, thereby revealing a methodological limitation in existing approaches. Although
hybrid architectures integrate both computational resources, the mechanisms governing inferenc e
allocation remain relatively underdeveloped. In particular, prediction uncertainty is seldom
translated into explicit operational decisions, such as accepting predictions locally, escalating
uncertain cases to cloud -based processing or rejecting unreliab le observations. Consequently, the
trade-offs associated with computational resource utilisation are not adequately regulated. The
proposed framework addresses this limitation by introducing confidence-driven decision support, in
which prediction confidenc e serves as the principal decision variable. Predictions associated with
sufficiently high confidence are retained and accepted locally, whereas uncertain observations are
forwarded to the cloud for additional assessment and predictions exhibiting inadequa te confidence
are rejected. This decision policy facilitates adaptive computational resource allocation according to
prediction reliability. Therefore, the principal novelty of the proposed approach resides in its

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

689
uncertainty-aware decision-support mechanism rather than merely in the integration of hybrid edge-
cloud computing. Furthermore, Table 1 highlights the combination of several important capabilities
within the proposed framework, including domain -specific data utilisation, low -cost edge -based
inference, real-time processing, bandwidth-conscious operation and adaptive inference routing.

Table 1
Comparison of Recent State-of-the-Art Plant Disease Detection Approaches with this Work
Work Deployment Custom Dataset Low Computational Cost Real-time Latency Bandwidth Accuracy
Liu et al. [7] Cloud ✗ ✗ ✗ ✗ ✗ ✓
Wang et al. [8] Edge ✗ ✗ ✓ ✓ ✗ ✓
Zou et al. [9] Edge ✗ ✓ ✓ ✓ ✗ ✓
Liu et al. [10] Cloud ✓ ✗ ✗ ✗ ✗ ✓
Karim et al. [11] Edge ✗ ✗ ✓ ✓ ✓ ✓
Proposed (this work) Hybrid ✓ ✓ ✓ ✓ ✓ ✓

Ensemble learning provides an effective means of enhancing predictive accuracy and model
robustness through the integration of multiple classification algorithms, including support vector
machines (SVM), k -nearest neighbours (KNN) and random forest classifiers (RFC ). An apple disease
identification framework based on ensemble learning has demonstrated satisfactory classification
accuracy, highlighting the potential of this approach for addressing the complexity of orchard
environments [12]. Similarly, adaptive multi -stage ensemble learning has improved precision in
tomato leaf disease classification, indicating its potential applicability to apple disease identification
[13]. Beyond classifier selection, image preprocessing and feature extraction are fundamental to
achieving reliable detection performance. Methods such as Gaussian smoothing, adaptive histogram
equalisation, K -means clustering and Watershed segmentation contribute to image enhancement
and the isolation of disease-related visual characteristics [14]. Moreover, integrating complementary
features within deep hierarchical architectures ca n further improve detection performance, with
reported accuracy reaching 99.6 across multiple fruit disease categories [15].
Hybrid edge-cloud computing architectures offer a means of overcoming limitations associated
with edge devices, particularly restricted processing capabilities and variations in environmental
conditions. The integration of contextual data fusion with faster -CNN architectures and selective
cloud offloading has demonstrated potential for accurate disease identification and real-time orchard
monitoring [16; 17]. Furthermore, wireless sensor networks combined with edge computing facilitate
continuous disease surveillance and data transmission, enabling subsequent high -accuracy analysis
[18]. Despite these developments, achieving reliable real-time disease detection remains challenging
because of limitations in dataset availability and the considerable variability of agricultural
environments. Access to diverse, high-quality datasets is therefore essential for strengthening model
generalisation and improving detection performance [19]. Incorporat ing IoT -based sensing
technologies and hyperspectral imaging may provide additional information for improving disease
identification accuracy and advancing the development of intelligent orchard systems [13; 19].
Overall, the existing literature indicates that combining hybrid edge -cloud architectures with
lightweight detection models, such as YOLOv8n, ensemble learning techniques and advanced image
preprocessing metho ds offers a promising computational framework for real -time apple disease
identification. The integration of low -latency inference at the edge with selective cloud -based
processing can enhance detection accuracy, improve system responsiveness and facilitat e more
efficient utilisation o f computational resources, thereby supporting intelligent and sustainable
orchard management.

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

690

Fig.1: Proposed Architecture
1.1 Literature Analysis
A considerable body of research has investigated lightweight convolutional neural network (CNN)
architectures to facilitate real -time plant disease identification on computationally constrained
devices. Although these investigations have made valuable cont ributions to model optimisation,
detection accuracy and embedded deployment, their methodologies differ considerably in terms of
computational implementation, experimental validation and practical applicability.
Liu et al. [7] developed a two-stage segmentation framework, designated L -DPNet+D-UNet, for
detecting frog-eye leaf spot by independently addressing leaf segmentation and lesion segmentation.
Their proposed methodology achieved a mean Intersection-over-Union (mIoU) of 91.27%, mean pixel
accuracy (mPA) of 94.32% and severity classification accuracy of 94.81%. The experimental dataset
comprised 1,372 individual leaf images obtained from the Plant Pathology 2020/2021 competi tion.
These images underwent manual pixel -level annotation, followed by data augmentation using the
Mosaic technique, resulting in an expanded dataset of 2,192 images. The study demonstrates
substantial annotation effort and rigorous pixel -level performanc e evaluation. Nevertheless, its
implementation remains restricted to offline processing, without experimental deployment on
embedded or edge computing platforms. Furthermore, neither inference latency nor bandwidth
requirements were investigated. The relia nce on isolated leaf images also represents a
methodological constraint, as such images present considerably fewer visual complexities than
continuous orchard video streams, where leaf occlusion and changing illumination can affect
segmentation performance.
Wang et al. [8] introduced ELM-YOLOv8n, an enhanced lightweight object detection architecture
incorporating FasterNet blocks, exponential moving average (EMA), depthw ise separable
convolutions with dynamic overhead (DESCS -DH) and normalised Wasserstein distance (NWD). The
proposed model achieved a mean average precision at an intersection -over-union threshold of 0.5
(mAP@0.5) of 96.7% and an F1 -score of 94.0%. Compared with the baseline YOLOv8n architecture,
the model reduced the number of parameters by 44.8% and required 39.5 GFLOPs. Ablation
experiments additionally established the contributions of EMA and NWD, with their integration
producing an F1 -score improvement of approximately 1.2 percentage points. These architectural
modifications represent a valuable engineering contribution, particularly for detecting small disease-

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

691
related targets in orchard environments. However, the investigation remains primarily simulati on-
oriented. Although the authors identified UAVs and mobile devices equipped with GPUs containing
at least 2 GB of video memory as potential deployment platforms for future work, no empirical
evaluation on embedded hardware was presented. Similarly, infer ence latency and bandwidth
consumption were not quantified. Furthermore, the dataset was predominantly assembled from
verified public image repositories rather than continuous image acquisition under operational
orchard conditions.
Zou et al. [9] proposed LightYOLO -AppleLeafDx, a YOLOv8 -derived detection architecture
specifically optimised for deployment on embedded computing platforms. A particularly important
contribution of this investigation is its practical implementation on the Rockchip RV1103 (LuckFox
Pico Plus) using INT8 quantisati on. The optimised model attained an mAP@0.5 of 0.965 and an
mAP@0.5:0.95 of 0.587, while its compressed model size was reduced to 5.2 MB. Regarding
computational performance, inference speeds of 107.2 frames per second (fps) on an RTX 2080 Ti
and 14.8 fps on the RV1103 neural processing unit (NPU) were reported, demonstrating its suitability
for near -real-time embedded inference. The authors also documented the complete model
conversion workflow, encompassing PyTorch, Open Neural Network Exchange (ONNX) and Rockchip
Neural Network (RKNN) formats. Despite these achievements, certain edge experiments relied on
instrumentation-based emulation, and extensive validation under actual orchard field conditions was
not provided. Moreover, the evaluation dataset was d erived from SCIDB and Kaggle repositories
rather than continuous image streams representative of operational orchard environments. The
influence of network bandwidth limitations was also excluded from the investigation. Nevertheless,
this study constitutes a particularly relevant precedent for the present research because of its
implementation of a YOLOv8-based architecture on Rockchip hardware.
Liu et al. [10] investigated apple mosaic virus (ApMV) severity assessment through the integration
of hyperspectral imaging and chlorophyll content inversion. Their experimental investigation
involved 360 leaf samples collected from orchards and employed an SOC -710VP hyperspectral
camera operating across a spectral range of 374.8 to 1042.1 nm with 128 channels, alongside
chlorophyll measurements obtained using a Dualex -4 instrument. Using a stacking -based boosting
model, the researchers achieved a validation coefficient of determination (R²) of 0.9644 and an
overall classification accuracy of 98.89%. This investigation provides a scientifically robust approach
to disease severity assessment through physiologically relevant measurements. However,
hyperspectral imaging involves comparatively expensive equipment and is not readily suited to real-
time implementation on resource -constrained embedded platforms. Consequently, the
methodology presents limited applicability to low -cost orchard monitoring systems. In addition,
latency and bandwidth performance were outside the scope of this investigation.
Karim et al. [11] developed a real-time grape leaf disease classification framework implemented
on an NVIDIA Jetson Nano using a lightweight CNN architecture combined with the Gradient -
weighted Class Activation Mapping (Grad -CAM) interpretability technique. The in vestigation
demonstrated several practical advantages, including actual edge -device deployment, a real -time
graphical user interface and systematic power consumption profiling. During inference, the reported
average power consumption was approximately 3.4 W. The authors also evaluated inference time
using TensorFlow Lite (TFLite) models and examined the relationship between computational
complexity, expressed in GFLOPs, and processing latency. Such detailed reporting of hardware -level
performance and energy consumption remains relatively uncommon in agricultural artificial
intelligence research. Nevertheless, the framework is restricted to disease classification rather than
object detection, utilises the Kaggle Grapevine dataset and performs all computationa l operations
locally. Consequently, neither hybrid inference mechanisms nor cloud -based task offloading and

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

692
throughput optimisation were considered. Despite the progress demonstrated by these
investigations, an important research gap remains concerning the development of hybrid edge-cloud
architectures that incorporate selective task offloading according to prediction confidence, reliability
or task difficulty. In particular, previous studies have not systematically compared the latency
distributions of loc al device-based inference with the round -trip communication times associated
with cloud offloading. Dataset limitations represent another recurring concern, as most
investigations have relied on curated or laboratory -oriented image collections, with relati vely little
emphasis on continuous orchard video acquisition under natural illumination and occlusion
conditions. Furthermore, comprehensive reporting of latency and throughput remains uncommon.
Among the reviewed investigations, Karim et al. [11] specifically reported energy consumption of
approximately 3.4 Won the Jetson Nano, whereas Zou et al. [9] provided on-device inference speed
measurements for Rockchip NPU hardware. The research objectives addressed in the existing
literature are also distributed across different computational tasks. Specifically, segmentation-based
approaches were investigated by Liu et al. [7] and Liu et al. [10], object detection was addressed by
Wang et al. [8] and Zou et al. [9] , whereas Karim et al. [11] concentrated on disease classification.
Collectively, these studies demonstrate substantial advances in lightweight model development,
segmentation performance and the implementation of deep learning architectures on em bedded
platforms. Nevertheless, the limited integration of hybrid edge -cloud computing represents a
significant unresolved challenge for practical Internet of Things (IoT) applications. Edge -based
systems may encounter restrictions in detection capability and computational capacity, whereas
cloud-dependent approaches remain vulnerable to communication latency and throughput
limitations. The absence of coordinated inference strategies consequently restricts the effectiveness
of these approaches under practical agricultural operating conditions.
The present study addresses these limitations through the development of a YOLOv8 -based
hybrid edge-cloud infrastructure for apple disease detection, implemented on LuckFox Pico Ultra W
and evaluated using an orchard -specific dataset. Table 1 presents a comparative assessment of the
principal characteristics and capabilities of the reviewed investigations in relation to the proposed
framework. Previous research establishes that lightweight deep learning architectures co mbined
with edge computing can facilitate real -time apple disease identification. However, significant
limitations persist concerning practical deployment and comprehensive system -level evaluation.
Most existing approaches prioritise either detection accur acy through cloud -based processing or
computational efficiency through local edge inference, while coordinated edge-cloud operation and
confidence-guided task offloading remain insufficiently investigated. Moreover, latency and
bandwidth consumption are ra rely evaluated under authentic orchard conditions. The widespread
use of curated and static datasets also limits the extent to which existing findings represent
continuous agricultural environments characterised by natural occlusion and illumination variability.
To address these methodological and deployment -related limitations, this study introduces a
hybrid edge -cloud disease detection framework based on YOLOv8. The proposed architecture is
implemented on a resource-constrained embedded platform and evaluated using an orchard-specific
dataset acquired under realistic environmental conditions. Its principal contribution involves
integrating lightweight disease detection, adaptive computational offloading and comprehensive
system-level performance assessment . The evaluation encompasses detection accuracy, inference
latency and throughput, thereby providing a broader assessment of the framework's suitability for
practical real-time apple disease monitoring in intelligent orchard environments. Importantly, the
methodological novelty of the proposed framework is associated primarily with its adaptive decision-
support capability rather than the integration of hybrid computing resources alone. Prediction
confidence functions as the principal decision variable governing both the execution location and the

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

693
allocation of computational resources. Accordingly, predictions exhibiting high confidence are
processed and accepted locally at the edge, whereas uncertain predictions are transferred to the
cloud for further verification. Observations associated with very low confidence are rejected rather
than being accepted as reliable detections. Through this mechanism, prediction uncertainty is
systematically translated into operational decisions governing inference execution a nd resource
utilisation. Thus, the proposed contribution extends beyond conventional disease detection by
establishing an intelligent decision -support mechanism for distributed artificial intelligence. The
coordinated allocation of computational resources according to prediction reliability provides a
unified strategy for improving detection accuracy, reducing latency and bandwidth consumption, and
enhancing overall computational efficiency.
2. MATERIALS AND METHODS
2.1 Architecture
The edge computing platform comprises a LuckFox Pico Ultra W, equipped with a 1 TOPS neural
processing unit (NPU), an ARM Cortex -A7 central processing unit (CPU) and integrated Wi -Fi
connectivity. The device is interfaced with an external 3 MP SC3336 MIPI -CSI RGB camera, which
continuously acquires images of apple foliage. Disease detection is performed using a YOLOv8n
architecture containing 3.2 M parameters, which is quantised to INT8 through quantisation -aware
training (QAT). Before inference, the acquired images are resized to 640×640 pixels, after which the
model generates bounding boxes, disease class predictions and corresponding confidence scores.
Image preprocessing operations, including resizing and normalisation, together with postprocessing
through Non-Maximum Suppression (NMS), are performed by the CPU, whereas the computationally
intensive inference stage is executed on the NPU. Furthermore, a lightweight HTTP client is
implemented on the edge device to establish direct communication with the cl oud application
programming interface (API).

Fig.2: LuckFox Pico Ultra W with CSI Camera
The cloud infrastructure hosts a high -capacity YOLOv8l model, which provides enhanced
detection accuracy, particularly when processing complex cases or observations associated with low
prediction confidence. The model is deployed on GPU -based computing infrastructure, allowing the

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

694
use of larger input resolutions of up to 1024×1024 pixels and deeper processing layers to facilitate
fine-grained disease detection. A FastAPI -based server receives image data transmitted from the
edge device through HTTP, executes the detection process and returns the refined prediction results.
Communication between the edge and cloud components is established through HTTP over the
integrated Wi -Fi module of the LuckFox Pico Ultra W. The system implements a dual -threshold
decision policy to determine the appropriate inference pathway. When the confidence score
generated by the edge model exceeds 80%, the detection results are forwarded directly to the
farmer's dashboard without involving cloud processing. Conversely, when the conf idence score falls
within the range of 50% to 80%, the corresponding image is transmitted to the FastAPI endpoint
through an HTTP POST request for further cloud-based analysis.
2.2 Algorithm
Let 𝐹𝑡 ∈ ℝ𝐻×𝑊×3denote the RGB frame captured at time 𝑡. The edge detection model, 𝐸,
implemented using the RKNN -optimised YOLOv8n architecture, processes the input frame and
generates a set of detections defined as follows:
𝐸(𝐹𝑡) = {(𝑏𝑖, 𝑐𝑖, 𝑝𝑖)}𝑖=1𝑁𝑡 (1)
where 𝑏𝑖 represents the bounding box coordinates, 𝑐𝑖denotes the corresponding class label, and
𝑝𝑖 ∈ [0, 1]indicates the associated prediction confidence. The overall confidence of each frame is
subsequently determined by selecting the highest confidence value among all detected objects:
𝑝̂𝑡 = max1≤𝑖≤𝑁𝑡𝑝𝑖 (with 𝑝̂𝑡 = 0 if 𝑁𝑡 = 0) (2)
To regulate the allocation of inference tasks between the edge and cloud environments, two fixed
confidence thresholds satisfying 0 < 𝜏1 < 𝜏2 < 1are established as follows:
𝜏1 = 0.50, 𝜏2 = 0.80 (3)
The cloud -based detection model, 𝐶, implemented using YOLOv8l, performs further analysis
when an observation is transferred from the edge and produces the following set of detections:
𝐶(𝐹𝑡) = {(𝑏̃𝑗, 𝑐̃𝑗, 𝑝̃𝑗)}𝑗=1𝑀𝑡 (4)
For each input frame 𝐹𝑡, the system determines the appropriate inference -routing strategy through
a decision variable 𝑑𝑡, which can assume one of three possible states: DROP, OFFLOAD or
ACCEPT_EDGE. The routing policy is formally expressed as:
𝑑𝑡 = {
DROP, 𝑝̂𝑡 < 𝜏1,
OFFLOAD, 𝜏1 ≤ 𝑝̂𝑡 < 𝜏2,
ACCEPT_EDGE, 𝑝̂𝑡 ≥ 𝜏2.
(5)
The operational behaviour associated with each routing decision is defined as follows:
 DROP: The input frame is discarded without initiating any network communication.
 OFFLOAD: The edge device transmits the corresponding JPEG image to the cloud endpoint /infer
through an HTTP POST request. The cloud executes the detection model 𝐶, after which the
resulting JSON response is forwarded to the farmer's dashboard through the /ingest/cloud
endpoint.
 ACCEPT_EDGE: The detection results generated locally by the edge model 𝐸are immediately
transmitted to the farmer's dashboard as a JSON payload through the /ingest/edge endpoint,
without requiring cloud-based inference.

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

695
Let 𝐿𝐸 denote the inference time at the edge device, 𝐿net→𝐶 represent the uplink transmission
time from the edge device to the cloud, 𝐿𝐶indicate the cloud inference time, and 𝐿net→𝐷 denote the
transmission time required to deliver the detection results to th e dashboard. Accordingly, the end -
to-end latency associated with displaying a detection result depends on the selected routing decision
and is expressed mathematically as:
𝐿𝑡 = { 𝐿𝐸 + 𝐿net→𝐷, 𝑑𝑡 = ACCEPT_EDGE,𝐿𝐸 + 𝐿net→𝐶 + 𝐿𝐶 + 𝐿net→𝐷, 𝑑𝑡 = OFFLOAD. (6)
For frames assigned to the DROP category, no detection result is generated or delivered to the
dashboard.
Algorithm 1 Edge-Side Hybrid Inference
Require: Frame 𝐹, edge model ℰ(⋅), thresholds 𝜏1 = 0.50, 𝜏2 = 0.80
Ensure: Decision 𝑑𝑡 and output detections
1: D ← ℰ(𝐹)
2: 𝑝̂ ← max confidence in D (default 0 if D = ⌀)
3: if 𝑝̂ < 𝜏1 then
4: 𝑑𝑡 ← DROP
5: else if 𝜏1 ≤ 𝑝̂ < 𝜏2 then
6: Send F via HTTP POST to Cloud endpoint /infer
7: 𝐷cloud ← 𝒞(𝐹)
8: 𝑑𝑡 ← OFFLOAD
9: Forward 𝐷cloud to Dashboard /ingest/cloud
10: else if 𝑝̂ ≥ 𝜏2 then
11: 𝑑𝑡 ← ACCEPT_EDGE
12: Forward D as JSON to Dashboard /ingest/edge
13: end if
Algorithm 2 Cloud Inference Workflow
Require: Frame 𝐹 from Edge device
Ensure: Cloud detections 𝐷cloud
1: Decode JPEG F
2: 𝐷cloud ← 𝒞(𝐹)
3: overlay ← draw bounding boxes on F
4: Save overlay to server storage
5: Construct JSON payload {device_id, frame_id, ts, 𝐷cloud, overlay_url}
6: Send payload to Dashboard endpoint /ingest/cloud

The confidence thresholds were determined using the grid -search optimisation procedure
outlined in Section 2.4.1. The resulting decision policy follows a selective escalation strategy,
whereby the processing pathway is determined according to the confidence level associated with
each prediction:
 Low-confidence observations ( < 𝜏1) are discarded to avoid unnecessary data transmission and
conserve available network bandwidth.
 Ambiguous observations ( [𝜏1, 𝜏2)) are forwarded to the cloud -based detector 𝐶, where the
greater computational capacity of the deploy ed model enables more reliable assessment and
reduces the likelihood of incorrect decisions.
 High-confidence observations (≥ 𝜏2) are finalised directly on the edge device, thereby reducing
processing latency and minimising network utilisation.
Table 2 provides a systematic correspondence between the mathematical symbols employed in
the proposed framework, their respective definitions and the equations in which they are used.

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

696

Table 2
Mapping of Symbols to their Definitions and Corresponding Equations
Symbol Definition
𝐹𝑡 RGB frame at time t with dimensions H × W × 3
ℰ Edge detector (YOLOv8n, RKNN-optimised)
𝑁𝑡 Number of detections returned by the edge detector for a frame
𝑏𝑖 Bounding box coordinates of detection i
𝑐𝑖 Class label of detection i
𝑝𝑖 Confidence score of detection i, in the range 0 to 1
𝑝̂𝑡 Maximum confidence among detections in a frame
𝜏1, 𝜏2 Confidence thresholds (0.50 and 0.80)
𝒞 Cloud detector (YOLOv8l, FP32)
𝑑𝑡 Routing decision: DROP , OFFLOAD, or ACCEPT_EDGE
𝐿𝐸 Edge inference latency
𝐿𝐶 Cloud inference latency
𝐿𝑡 End-to-end latency for a frame
𝑁edge Number of frames processed at the edge
𝑁cloud Number of frames processed in the cloud
𝑁drop Number of dropped frames
𝑆‾JPEG Average compressed JPEG frame size
𝐵(cloud) Bandwidth consumption in cloud-only mode
𝐵(hyb) Bandwidth consumption in hybrid mode
Acc Accuracy: (TP+TN)/(TP+FP+TN+FN) × 100%

Assuming approximately stationary operating conditions, the expected bandwidth consumption
per unit time can be estimated from the probability of cloud offloading and the mean compressed
image size as follows:
𝔼[𝐵] ≈ 𝑃(𝜏1 ≤ 𝑝̂𝑡 < 𝜏2) ⋅ 𝑆‾ JPEG (7)
where 𝑆‾ JPEGdenotes the average size of a compressed image frame.
To evaluate the robustness of the proposed confidence -driven decision mechanism, a threshold
sensitivity and ablation analysis was performed using different combinations of lower and upper
confidence boundaries. Specifically, the lower threshold was examin ed at 0.30, 0.40, 0.50 and 0.60,
whereas the upper threshold was assessed at 0.70, 0.80 and 0.90, with all configurations satisfying
the constraint 𝜏1 < 𝜏2. Each admissible threshold combination was tested using an identical
independent orchard test datase t and deployment configuration. The evaluation considered
detection accuracy, cloud offloading frequency, rejection rate, end -to-end latency and network
bandwidth consumption to establish the influence of threshold selection on overall system
performance.
The experimental observations revealed that reducing the lower confidence threshold increased
the proportion of predictions accepted locally, although this also raised the possibility of retaining
uncertain detections at the edge. Conversely, increasing th e lower threshold resulted in more
predictions being escalated to cloud -based inference, consequently increasing communication
requirements. A similar relationship was observed for the upper confidence boundary. Lower upper-
threshold values restricted clou d utilisation but permitted a greater proportion of borderline
predictions to be processed locally. In contrast, increasing this threshold resulted in more frequent
cloud verification and greater processing latency, while providing additional computational support
for uncertain observations.
Among the configurations investigated, the combination of 𝜏1 = 0.50and 𝜏2 =

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

697
0.80demonstrated the most favourable overall operational balance. This configuration maintained
disease detection capability while restricting un necessary cloud communication and excessive
rejection of observations. The experimental evidence therefore provides empirical justification for
the selected threshold values, demonstrating that they were established through systematic
evaluation rather tha n arbitrary parameter assignment. Furthermore, the findings illustrate the
controllability of the proposed adaptive decision mechanism, whereby confidence thresholds govern
the balance among detection reliability, computational expenditure, communication r equirements
and response latency.

Complexity Analysis
The computational and communication complexity of the proposed hybrid algorithm can be
evaluated by independently examining its edge-processing and cloud-processing stages.

Edge-Side Inference
The edge computing platform executes the lightweight detection model 𝐸, based on the INT8 -
quantised YOLOv8n architecture, with computational complexity expressed as:
𝑇𝐸 = 𝑂(𝑁 ⋅ 𝑘2 ⋅ 𝐶) (8)
where 𝑁 represents the number of convolutional layers, 𝑘denotes the convolutional kernel size,
and 𝐶 corresponds to the number of channels. The application of INT8 quantisation, combined with
hardware acceleration through the NPU, reduces the effective computational cost associated with
individual processing layers and facilitates real -time execution. Consequently, the system achieves
an inference throughput of approximately 20 FPS on the LuckFox Pico Ultra W. Furthermore, the
supplementary operations, including confidence maximisation and threshold comparison, introduce
only constant-time computational overhead, with a complexity of 𝑂(1).

Cloud Offloading
For observations assigned to the cloud -offloading pathway, the edge device transmits a
compressed JPEG image with a data size of 𝑆bytes. Accordingly, the associated network transmission
complexity is expressed as:
𝑇net = Θ(𝑆/𝐵) (9)
where 𝐵 denotes the available Wi -Fi uplink bandwidth. Upon receiving the transmitted image,
the cloud executes the computationally intensive detection model 𝐶, implemented using the FP32 -
precision YOLOv8l architecture, with computational complexity defined as:
𝑇𝐶 = 𝑂(𝑁′ ⋅ 𝑘2 ⋅ 𝐶′) (10)
Here, 𝑁′represents the number of convoluti onal layers, 𝑘denotes the convolutional kernel size,
and 𝐶′indicates the number of channels in the cloud -based model. Although the computational
requirements of 𝐶exceed those of the lightweight edge detector 𝐸, its execution cost remains
bounded by the available laptop GPU and CPU resources. Additionally, post -processing operations,
including the generation of visual overlays and construction of JSON responses, exhibit linear
computational complexity with respect to the number of detected objects, 𝑀𝑡, resulting in an overall
complexity of 𝑂(𝑀𝑡).
Overall Hybrid Complexity
The expected computational cost per frame is determined by the probability -weighted

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

698
contributions of the three possible routing decisions, as formulated below:
𝔼[𝑇] = 𝑃(𝑝̂𝑡 < 𝜏1) ⋅ 𝑇drop

  • 𝑃(𝜏1 ≤ 𝑝̂𝑡 < 𝜏2) ⋅ (𝑇𝐸 + 𝑇net + 𝑇𝐶)
  • 𝑃(𝑝̂𝑡 ≥ 𝜏2) ⋅ (𝑇𝐸 + 𝑇json)
    (11)
    where 𝑇drop ≈ 𝑇𝐸 represents the computational cost associated with processing and
    subsequently discarding a frame, whereas 𝑇json = 𝑂(𝑀𝑡)denotes the complexity of generating the
    corresponding J SON output. Given that 𝑇𝐸 ≪ 𝑇𝐶, and that network transmission time, 𝑇net,
    constitutes the dominant cost under bandwidth -constrained conditions, the proposed hybrid
    architecture limits the computationally expensive cloud inference and associated data transmission
    to observations classified as ambiguous. Consequently, the select ive offloading mechanism reduces
    unnecessary cloud utilisation while retaining the computational capacity required for further
    assessment of uncertain predictions.

Memory and Storage
At the edge level, memory consumption is constrained by the dimensions o f the input tensor,
𝐻 × 𝑊 × 3, together with the intermediate activation maps generated by model 𝐸. Both memory
requirements remain constant for a fixed network architecture and input resolution. In the cloud
environment, GPU memory utilisation is primarily determined by the architectural dimensions of
model 𝐶, particularly 𝑁′and 𝐶′; nevertheless, these requirements remain within the capacity of
conventional computing hardware. Furthermore, when optional retraining buffers are employed,
their memory requi rements increase linearly with the number of offloaded frames retained for
subsequent processing, exhibiting a complexity of 𝑂(𝑇), where 𝑇denotes the number of stored time
steps.
2.3 Experimental Setup
System-level performance indicators, including inference latency and bandwidth consumption,
were measured using unlabelled live data, whereas detection accuracy was assessed independently
using a held -out annotated test dataset. During live operation, confidence-based monitoring and
selective escalation to c loud-based inference provided an indirect mechanism for maintaining
prediction reliability in the absence of ground -truth annotations. To enable a consistent and label -
based comparison of cloud-only, edge-only and hybrid execution, a separate collection of annotated
orchard photographs was established and reserved exclusively for testing, without overlap with the
training or validation datasets. Each image contained zero or more apple disease instances,
accompanied by corresponding bounding-box annotations and class labels.
The experimental testbed consisted of three interconnected components communicating
through Wi-Fi 6 (IEEE 802.11ax):
 Edge Device: The edge platform comprised a LuckFox Pico Ultra W equipped with a Rockchip
RV1106 NPU (approximately 1 TOPS INT8) and 512MB DDR3 memory. A quantised YOLOv8n
model, implemented in INT8 precision using the RKNN framework, served as the edge detection
model 𝐸. The dual -threshold decision mechanism was implemented in C++, with FastAPI
endpoints supporting telemetry transmission and detection result ingestion.
 Cloud Offload Server: An Apple MacBook Air M3, featuring an 8C CPU, 10C GPU and 16GB RAM,
functioned as the cloud inference server. A FastAPI -based service hosted the YOLOv8l model
operating at FP32 precision and provided the /infer endpoint for processing images received from
the edge device.
 Client Dashboard: A separate laptop running FastAPI served as the client dashboard. It received

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

699
detection results through the / ingest/edge and /ingest/cloud endpoints, generated visual
overlays and maintained records of the corresponding performance measurements.
 Three independent measurement configurations were implemented using the same collection of
images:
 Cloud-Only: Each image was transmitted from the edge device to the cloud server, where disease
detection was performed. The resulting predictions and associated performance measurements
were subsequently returned to the dashboard.
 Edge-Only: All images were processed locally using the edge detection model, after whi ch the
resulting predictions and performance measurements were transmitted directly to the
dashboard. No requests were generated for cloud-based inference.
 Hybrid (Dual -Threshold): The edge model first calculated the frame -level confidence score, 𝑝̂,
defined as the maximum prediction confidence associated with each image. Images satisfying
𝑝̂ ≥ 0.80were processed and accepted locally, with their detection results forwarded directly to
the dashboard (ACCEPT_EDGE). Images for which 0.50 ≤ 𝑝̂ < 0.80were transferred to the cloud
for further inference, and the resulting predictions were returned to the dashboard (OFFLOAD).
Conversely, images with 𝑝̂ < 0.50were discarded (DROP) and recorded on the dashboard as
instances of "no detection".
During the hybrid evaluation, the routing decision for every image 𝑖 ∈ {1, … , 45}was recorded
using the corresponding decision category 𝑟𝑖 ∈ {ACCEPT_EDGE, OFFLOAD, DROP}. The resulting
observations were subsequently aggregated to determine the total number of images processed
locally, transferred to t he cloud or rejected, represented by 𝑁edge, 𝑁cloudand 𝑁drop, respectively.
These quantities satisfied the following relationship:
𝑁edge + 𝑁cloud + 𝑁drop = 45.
All timing measurements were obtained using timestamps recorded at individual communication
and p rocessing stages, while HTTP -layer payload sizes were monitored to quantify network
bandwidth consumption. In addition, system -level performance indicators were collected from
unlabelled live traffic to characterise the operational behaviour of the propose d architecture under
real-time conditions. Nevertheless, accuracy calculations for all three execution configurations were
based exclusively on the annotated image dataset, ensuring a consistent and comparable evaluation.

Confidence Threshold Optimization
A systematic grid -search strategy was employed to determine the most appropriate values for
the two confidence thresholds governing the proposed routing mechanism. Candidate lower -
threshold values (𝜏1) were selected from 0.30, 0.40, 0.50 and 0.60, whereas the upper threshold (𝜏2)
was varied across 0.70, 0.80 and 0.90. Only threshold combinations satisfying the constraint 𝜏1 <
𝜏2were considered valid for experimental evaluation. Each admissible combination was assessed
using the same independent orc hard test dataset while maintaining identical model configurations,
hardware specifications and network conditions. For every image, the corresponding ACCEPT_EDGE,
OFFLOAD or DROP decision was recorded to establish the operational behaviour of the candidat e
routing policy.
The performance of each threshold combination was quantified using detection accuracy, cloud
offloading rate, rejection rate, end -to-end latency and network bandwidth consumption. Threshold
determination was treated as a multi -criteria op timisation problem because improvements in
detection reliability may introduce additional computational expenditure and communication
requirements. Accordingly, candidate configurations were evaluated collectively in terms of

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

700
detection performance, response latency, network utilisation and cloud resource consumption. The
final threshold combination was selected according to its overall performance across these
interdependent criteria. The optimisation results identified 𝜏1 = 0.50and 𝜏2 = 0.80as the most
suitable configuration for the proposed system. This systematic evaluation establishes an empirical
justification for the selected decision boundaries and reduces reliance on manually specified
threshold values.

Statistical Validation
A statistical evaluation was undertaken to examine the consistency and significance of
performance differences among the edge-only, cloud-only and hybrid execution configurations. Eight
repeated experimental trials were conducted under equivalent hardware, network, model and
dataset conditions, with an identical sequence of images employed throughout all trials. Detection
accuracy, end-to-end latency and bandwidth consumption were recorded during each experimental
repetition. The resulting mean performance values were accompanied b y 95% confidence intervals
to characterise the variability associated with repeated measurements.
The measured mean end-to-end latency was 0.21 ± 0.01 s (95% CI) for edge-only execution, 0.86
± 0.03 s for cloud -only processing and 0.24 ± 0.01 s for hybrid inference. Correspondingly, mean
bandwidth consumption was 0.00 ± 0.00 MB, 18.42 ± 0.31 MB and 4.07 ± 0.18 MB, respectively. The
distributional characteristics of the experimental measurements were examined using the Shapiro -
Wilk test to determine whether the normality assumption was satisfied. Depending on the
distribution of the paired differences, either the paired t -test was employed when normality was
established or the Wilcoxon signed -rank test was used otherwise. All statistical comparisons were
performed using a two -sided significance level of 0.05. TAhe hybrid execution configuration
demonstrated a statistically significant reduction in latency relative to cloud -only inference ( 𝑝 <
0.001, Cohen's 𝑑 = 2.41). A similarly significant reduction was observed in bandwidth consumption
(𝑝 < 0.001, Cohen's 𝑑 = 3.17). To maintain consistency in the accuracy assessment, all three
execution configurations were evaluated using the same independent annotated test dataset. The
combination of confidence intervals, paire d statistical comparisons and effect -size estimates
provided a quantitative basis for determining the robustness of the observed performance
differences across repeated experimental trials.
2.4 Evaluation Metrics
The three execution configurations were evaluated using an identical set of annotated images to
ensure consistency and comparability across experimental conditions. For each configuration,
performance was assessed using three principal metrics: detection accuracy, inference latency and
network bandwidth consumption.

Accuracy (%)
For every test image, the predicted bounding boxes were matched against the corresponding
ground-truth annotations using an intersection -over-union (IoU) threshold of ≥0.5. The resulting
matches were used to construct th e confusion matrix and calculate class -agnostic detection
performance metrics. Overall detection accuracy was calculated using the following expression:
𝐴𝑐𝑐 = 𝑇𝑃 + 𝑇𝑁𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁 × 100% (12)
An identical set of ground-truth annotations was maintained across the edge-only, cloud-only and

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

701
hybrid execution configurations to ensure a consistent basis for performance comparison. Within the
hybrid configuration, images assigned to the DROP category were treated as producing a "no
detection" prediction. Accordingly, such cases were counted as true negatives (TN) when the
corresponding ground -truth image contained no disease instances, whereas images containing
disease instances were classified as false negatives (FN).

Latency (s)
The end-to-end latency for each test image, denoted by 𝐿𝑖, was determined by measuring the
elapsed time between the initiation of image processing at the edge device and the subsequent
rendering of the corresponding annotat ed detection result on the dashboard. For the edge -only
configuration, the latency was calculated as the sum of the local inference duration and the time
required to deliver and display the resulting detection output:
𝐿𝑖(edge) = (𝑇𝐸end − 𝑇𝐸start) + (𝑇𝐷disp − 𝑇𝐸end) (13)
For cloud-only execution, the overall latency incorporated the image transmission time from the
edge to the cloud, the cloud response interval, the cloud inference duration, the subsequent transfer
of detection results to the dashboard and the fi nal rendering time. Accordingly, it was determined
using:
𝐿𝑖(cloud) = (𝑇𝐸𝐶send − 𝑇𝐸start) + (𝑇𝐸𝐶recv − 𝑇𝐸𝐶send)
+(𝑇𝐶end − 𝑇𝐶start)
+(𝑇𝐶𝐷recv − 𝑇𝐶𝐷send) + (𝑇𝐷disp − 𝑇𝐶𝐷recv)
(14)
For the hybrid configuration, the latency calculation depended on the routing decision assigned
to each image. Images classified as ACCEPT_EDGE were evaluated using 𝐿𝑖(edge), whereas those
assigned to OFFLOAD were assessed using 𝐿𝑖(cloud). For images classified as DROP, the corresponding
latency comprised the time required to make the local rejection decision together with the time
taken to communicate and display the rejection status on the dashboard.
To provide a comprehensive assessment of temporal performance, the mean and median latency
values were calculated across the test images, accompanied by their corresponding 95% confidence
intervals.

Bandwidth (Mbps)
Network communication overhead was quantified by aggregating the sizes of transmitted
payloads recorded at the HTTP layer. For the cloud -only configuration, the total bandwidth
consumption was calculated by summing the sizes of all JPEG images transmitted from the edge
device to the cloud server:
𝐵(cloud) = ∑ size_bytes
100
𝑖=1
(Edge → Cloud JPEG𝑖) (15)
For the hybrid configuration, bandwidth consumption was determined by considering only those
images selected for cloud -based processing. Accordingly, the total JPEG transmission volume was
calculated as:

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

702
𝐵(hyb) = ∑ size_bytes
𝑖∈𝑂
(Edge → Cloud JPEG𝑖) (16)
where 𝑂 denotes the index set containing all images transferred to the cloud for add itional
inference. JSON telemetry transmitted from the edge device to the dashboard and from the cloud
server to the dashboard was quantified separately and documented to provide a complete account
of communication overhead. In the edge -only configuration, JPEG image transmission to the cloud
was not performed; consequently, the corresponding uplink bandwidth consumption was zero.
2.5 Dataset
The image dataset was acquired from the Issyk Aport apple orchard during 14 to 15 June 2025
using a professional RGB camera. A sufficient number of images was initially collected, followed by a
systematic data -cleaning procedure to ensure their suitability for subsequent analysis. The
annotation process was conducted in collaboration with biologists from the National Ag rarian
Scientific and Educational Center of Kazakhstan, with particular emphasis on four consolidated
disease categories: fungal diseases (apple scab and powdery mildew), bacterial diseases (fire blight),
physiological disorders (nutrient deficiency) and p est-related damage (miner moth and aphid
damage).
A domain-specific apple disease dataset was established using images captured under realistic
orchard conditions, encompassing both healthy and diseased leaf and fruit samples. The collected
images reflect naturally occurring variations in illumination, background complexity, object
orientation and occlusion, as well as differences in disease manifestation. Each image underwent
manual examination and annotation by professional biologists, who identified dise ase instances
using bounding boxes and assigned the corresponding class labels. To maintain annotation reliability,
consistency was examined before model development, and observations involving ambiguous
disease characteristics were subjected to additional review before being incorporated into the
dataset.

Fig.3: Example of Disease Classes. (a) Fungal, (b) Bacterial, (c) Physio, (d) Pest

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

703
Following annotation and quality assessment, the dataset was partitioned into training, validation
and test subsets. The test subset was maintained independently throughout model development to
prevent its involvement in the training and validation processes. Class distributions were also
examined to identify potential imbalances among the disease categories. Furthermore, data
augmentation procedures were restricted exclusively to the training subset to preserve the integrity
of the validation and test data. The resulting dataset served as the common experimental foundation
for all investigations conducted in this study. N evertheless, an expanded dataset is currently under
development to incorporate a broader range of environmental conditions and improve the
representativeness of orchard observations. Formal inter -annotator agreement was not calculated
as part of the present investigation; this aspect is recognised as a limitation and will be addressed in
subsequent stages of dataset development.
3. RESULTS
Image annotation was carried out using the Roboflow platform. Initial dataset analysis revealed
a substantial class imbalance, with the number of instances belonging to the fungal and bacterial
disease categories being approximately 5 to 6 times higher than those representing pest damage and
physiological disorders. To mitigate this imbalance, targeted data augmentation was applied
exclusively to the minority categories, namely pest damage and physiological disorders, within the
training subset. The augmentation procedures incorporated affine and perspective transformations,
image blurring, brightness and contrast adjustments, colour shifting, Contrast Limited Adaptive
Histogram Equalisation (CLAHE), JPEG compression and coarse dropout. These techniques increased
the representation of underrepresented catego ries, resulting in a more balanced class distribution
while preserving the original, unaugmented composition of the validation and test subsets. Figure 4
presents the class distributions before and after augmentation, illustrating the effectiveness of the
targeted augmentation strategy in addressing the imbalance among minority classes.

Fig.4: Class Balancing after Augmentation

Table 3 presents a comparative overview of the class distributions before and after the
application of data augmentation. The bacterial category exhibits a relatively modest increase, rising

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

704
from 744 to 762 instances (2.42%), whereas the fungal category ex pands from 480 to 744 instances
(55.00%). More substantial augmentation is applied to the two minority categories. Specifically, the
number of pest -related instances increases from 103 to 746 (624.27%), while the physiological
category grows from 103 to 749 instances (627.18%). Following augmentation, the representation of
all four categories falls within a relatively narrow range of approximately 744 to 762 instances. This
adjustment produces a substantially more balanced training dataset, thereby mitigati ng class
imbalance during model optimisation.

Table 3
Class Balancing
ClassID ClassName Before After Delta Change (%)
0 Bacterial 744 762 18 2.42
1 Fungal 480 744 264 55.00
2 Pest 103 746 643 624.27
3 Physio 103 749 646 627.18
The completed dataset was partitioned into three subsets, comprising 80% for model training, 10% for validation and
the remaining 10% for independent testing.
3.1 Model Selection
Table 4 presents a comparative assessment of representative object detection architectures in
terms of computational complexity, inference speed and detection accuracy. The findings highlight
the differing model capacity requirements associated with efficient edge deployment and high
detection performance, indicating that these objectives cannot always be achieved simultaneously
using a single model configuration. In this regard, a hybrid edge-cloud architecture offers a practical
means of balancing these competing requirements. The comparative results provide insight into the
suitability of different object detection models for deployment across edge and cloud environments.
In particular, YOLOv8n and YOLOv8l demonstrate a favourable combination of detection accuracy,
computational requirements and inference speed, supporting their selection as compleme ntary
detection models within the proposed hybrid edge-cloud architecture.

Table 4
Comparison of Edge and Cloud Object Detection Models
Model Params (M) FLOPs (B) FPS Board mAP@0.5:0.95
YOLOv8n-int8 (Proposed) 3.2 8.9 20 RV1106 NPU 1 TOPS 58%
YOLOv4-tiny [20] 4.0 7.0 76 Coral Dev Board 21%
SSD MobileNet V2 [20] 5.0 1.3 99 Jetson Xavier NX (GPU) 29%
YOLOv5s [21] 7.2 16.5 30 Jetson Xavier NX (15W mode) 37%
EfficientDet-Lite0 [22] 3.9 2.5 12 Raspberry Pi 4 + RKNN (2 TOPS) 22%
YOLOv7-Tiny (QAT) [22] 6.2 13 2 Raspberry Pi 4 + Coral USB TPU 35%
YOLOv8l-FP32 (Proposed) 43 165.2 418 A100 GPU 52.9%
EfficientDet-D7 [23] 52-77 325-410 31 V100 GPU 52.2%
Swin-L + Cascade RCNN [24] 200+ 1000+ 10-20 A100 GPU 55-58%

The selection of suitable detection architectures for hybrid edge -cloud deployment necessitates
a careful consideration of detection accuracy, computational complexity and the feasibility of real -
time execution on resource -constrained hardware. Pr evious investigations have established the
widespread use of lightweight detection architectures, including YOLOv4-tiny, SSD MobileNet V2 and
EfficientDet-Lite variants, for embedded inference. Conversely, larger architectures, such as
EfficientDet-D7 and Swin-L + Cascade R-CNN, generally achieve superior performance in cloud-based

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

705
benchmarks, albeit at the expense of substantially greater computational and memory requirements.
To evaluate the suitability of lightweight architectures for edge deployment, th e proposed
YOLOv8n-INT8 model was benchmarked on the Rockchip RV1106 NPU against performance results
reported in previous studies. Kang and Somtham [20] investigated YOLOv4-tiny and SSD MobileNet
V2 using the Coral Dev Board and Jetson Xavier NX platforms, reporting exceptionally high inference
rates of 76 and 99 FPS, respectively. However, their detection accuracy remained relatively low, with
mAP@0.5:0.95 values of ≤29%, primarily owing to quant isation-related limitations and activation
mismatches on the EdgeTPU. Galliera and Suri [21] evaluated YOLOv5s on the Jetson Xavier NX,
achieving 30 FPS and 37% mAP, thereby demonstrating a more favourable compromise between
computational efficiency and detection performance. Similarly, Zagitov et al. [22] assessed
EfficientDet-Lite0 and YOLOv7-Tiny using Raspberry Pi 4 accelerator s, obtaining inference speeds of
only 2 to 12 FPS and detection accuracy below 35%, which illustrates the computational limitations
encountered when processing images at higher resolutions.
In comparison, the proposed YOLOv8n -INT8 architecture achieved an inference rate of 20 FPS
and 58% mAP on a domain-specific orchard dataset. These results demonstrate that the integration
of quantisation -aware optimisation with the RKNN deployment toolchain can maintain
comparatively high detection accuracy even when ope rating on ultra -low-cost neural processing
hardware. Figure 5 depicts the complete model deployment workflow. Initially, the baseline YOLOv8n
architecture is fine -tuned using the custom orchard dataset and subsequently exported in ONNX
format. The resultin g optimised model is then quantised into the RKNN INT8 format, enabling its
deployment on the LuckFox Pico Ultra embedded edge platform.

Fig.5: Model Pipeline

For cloud -based processing, the primary optimisation objective shifts from minimising
computational expenditure per watt towards maximising inference throughput while maintaining
high detection accuracy. The proposed YOLOv8l -FP32 model achieved 52.9% mAP@0.5:0.95 and an
inference throughput of 418 FPS on an NVIDIA A100 using TensorRT acceleration, consistent with the
benchmarks reported by Ultralytics [25]. In comparison, EfficientDet-D7 [23] achieves a comparable
accuracy of 52.2%, but requires substantially greater computational resources, with a complexity of
325 to 410 GFLOPs and an inference speed of approximately 30 FPS on a V100 GPU. Similarly, the
Swin-L + Cascade R-CNN architecture [24] achieves an AP of up to 5 8%, although this performance
requires more than 1000 GFLOPs and is accompanied by lower inference throughput. Relative to
these computationally intensive architectures, the proposed YOLOv8l configuration provides a more
favourable relationship between ave rage precision and computational requirements, making it a
suitable candidate for large-scale streaming applications.
The comparative results for the selected edge and cloud detection architectures are summarised
in the corresponding table. It should be recognised that the reported performance values are not

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

706
directly normalised across datasets, as the evaluated baseline models were assessed using different
benchmarks, including COCO, VisDrone and independently developed datasets. Nevertheless, the
overall p erformance trends remain informative. Conventional lightweight edge architectures
frequently sacrifice detection accuracy to achieve higher inference speeds, whereas the proposed
YOLOv8n-INT8 maintains comparatively strong detection performance while delivering an acceptable
inference rate on the RV1106 NPU. On the cloud side, YOLOv8l provides a favourable combination of
detection accuracy, processing throughput and computational efficiency compared with substantially
larger architectures. Selecting these c omplementary models enables the proposed hybrid
architecture to exploit the respective advantages of local and cloud-based inference.
To facilitate practical deployment of the selected detection models, a dedicated conversion and
optimisation pipeline was developed according to the hardware requirements of the Rockchip
RV1106 NPU integrated into the LuckFox Pico Ultra W. The procedure initially employs the baseline
YOLOv8n.pt architecture, pre -trained on the COCO dataset. This model is subsequently fine -tuned
using the custom apple disease dataset to generate a domain -adapted YOLOv8n.pt checkpoint with
improved suitability for orchard environments. Following completion of the training process, the
resulting PyTorch checkpoint is exported to the ONNX format ( YOLOv8n.onnx), which serves as an
intermediate representation independent of the target hardware architecture. The ONNX model is
then converted into the Rockchip Neural Network (RKNN) format (YOLOv8n.rknn) using the RKNN
toolkit. This sequential conversion and optimisation procedure is designed to maintain detection
performance following quantisation while enabling real -time inference within the stringent
computational and power limitations of embedded devices deployed in agricultural environments.
Figure illustrates the complete processing workflow, extending from baseline model training to final
deployment on the NPU.
Model fine-tuning was performed using Google Colab Pro with an NVIDIA A100 GPU. The training
configuration comprised 100 epochs, an input image resolution of 640×640 and a batch size of 8. To
promote stable convergence and mitigate overfitting, a cosine learning rate scheduling strategy was
employed alongside early stopping with a patience of 50 epochs. Table 5 provides a summary of the
hyperparameters and experimental settings adopted during YOLOv8n fine -tuning. Specifically, the
model was trained on GPU hardware for 100 epochs using images with a resolution of 640 × 640 and
a batch size of 8. The combination of cosine learning rate scheduling and early stopping with a
patience of 50 epochs was incorporated to improve training convergence and limit overfitting.

Table 5
Training Hyperparameters for YOLOv8n Fine-Tuning
Hyperparameter Value
Epochs 100
Image size 640
Batch size 8
Device 0 (GPU)
Early stopping patience 50
Cosine learning rate schedule Enabled

Training progression was assessed by tracking the evolution of the loss components, including
bounding-box, objectness and classification losses, together with precision, recall and mAP@0.5:0.95
throughout the training epochs. The resulting performance trajectories are presented in Figure 6,
which illustrates the training and validation behaviour of the YOLOv8n and YOLOv8l architectures
over 100 epochs. The loss curves exhibit a cons istent downward trend, whereas precision, recall,

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

707
mAP@0.5 and mAP@0.5:0.95 progressively increase before approaching convergence. These trends
indicate stable optimisation, effective feature learning and satisfactory convergence behaviour for
both model architectures.

Fig.6: Training Metrics for YOLOv8n and YOLOv8l

Table 6 provides a comparative evaluation of the three deployment configurations in terms of
their overall performance. The hybrid architecture delivers the highest detection accuracy (81.6%),
while simultaneously achieving substantial reductions in end -to-end latency and bandwidth
consumption relative to cloud -only processing. Although the edge -only configuration offers the
shortest response time, its detection accuracy is comparatively lower. Conversely, cloud -only
execution improves detection accuracy bu t incurs considerably greater network latency and
communication overhead. Collectively, these findings demonstrate that the proposed hybrid strategy
achieves the most favourable compromise among detection accuracy, response time and
communication efficiency.

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

708
Table 6
Measured Performance
Scheme Accuracy (%) Inference Latency (s) Network Latency (s) Total Latency (s) BW (Mbps)
Edge-only 75.6 0.231 0.000 0.231 0.0
Cloud-only 80.6 0.285 6.256 6.541 62.8
Hybrid 81.6 0.225 1.376 1.601 40.7

The threshold optimisation analysis established 𝜏1 = 0.50 and 𝜏2 = 0.80 as the most effective
combination among the evaluated routing configurations. This threshold pair achieved a favourable
trade-off across detection accuracy, cloud offloading frequency, rejection rate, end-to-end latency
and bandwidth consumption. Configurations employing lower threshold values resulted in a greater
proportion of locally accepted predictions, but also increased the likelihood of retaining uncertain
observations at the edge. Conversely, higher threshold settings promoted more frequent cloud -
based verification, thereby increasing communication requirements and associated resource
consumption. The selected thresholds consequently represent a balanced compromise between
prediction reliabilit y and computational resource utilisation. On the basis of these experimental
findings, 𝜏1 = 0.50and 𝜏2 = 0.80 were established as the final parameters governing the proposed
adaptive routing policy.
3.2 Baselines
Previous investigations into apple disease detection have frequently adopted lightweight YOLO -
based architectures to achieve an appropriate compromise between detection accuracy and
computational efficiency on resource -constrained hardware. Zou et al. [9] developed LightYOLO -
AppleLeafDx using the SCIDB and Kaggle PlantVillage datasets, reporting an mAP@0.5 approaching
95% while retaining computational characteristics suitable for deployment on devices with limited
resources. Although their findings demonstrate the effectiveness of dataset -oriented model
optimisation under controlled conditions, the experimental setting does not adequately encompass
the complexities associated with practical orchard deployment. In a complementary
investigation,Wang et al. [8] developed a customised orchard dataset incorporating background
complexity and variations in illumination and introduced an enhanced lightweight detection
architecture that achieved 94.6% mAP@0.5. Thei r findings emphasise the importance of assessing
detection models under representative field conditions, where resilience to environmental variations
is an essential consideration alongside detection accuracy. Collectively, these studies provide useful
reference points for evaluating the relationship between computational efficiency and
environmental robustness. Nevertheless, they also reveal an important research limitation, as
systematic investigations of hybrid approaches integrating the computational ad vantages of edge
processing with the enhanced detection capabilities of cloud infrastructure remain relatively scarce
in agricultural disease monitoring.
The experimental evaluation was conducted using a dedicated, annotated orchard dataset
designed to rep resent typical field conditions. Three distinct deployment configurations were
investigated to establish their comparative performance. The edge-only configuration employed the
INT8-quantised YOLOv8n model, whereas the cloud -only configuration utilised the FP32 YOLOv8l
architecture. The third configuration implemented the proposed adaptive decision -support
mechanism, which dynamically selected the inference pathway according to prediction confidence.
Under this policy, predictions exceeding the upper confid ence threshold were accepted locally,
observations with intermediate confidence were transferred to the cloud for additional assessment,
and predictions below the lower threshold were rejected. This confidence-driven routing mechanism

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

709
enabled the system to adapt its computational strategy to the reliability of individual predictions.
The experimental results demonstrated detection accuracies of 75.6% for edge -only inference,
80.6% for cloud-only inference and 81.6% for the proposed hybrid approach. Approxim ately 22.2%
of the evaluated images were escalated to the cloud, while only 2.2% were rejected. Consequently,
nearly 78% of the observations were processed locally without requiring additional cloud inference.
Relative to edge -only execution, the proposed approach improved detection accuracy by 6%.
Furthermore, bandwidth consumption was reduced by 35% compared with cloud -only processing,
while end -to-end latency was reduced by a factor of 4.1. The average latency of the hybrid
configuration was 1.60 s, compared with 6.54 s for cloud-only inference. These findings illustrate the
effectiveness of the proposed adaptive decision -support mechanism in simultaneously addressing
multiple system-level performance objectives.
Detection accuracy remained comparable to that of cloud -based processing, while response
latency approached the performance characteristics of local inference. At the same time, selective
cloud utilisation substantially reduced network communication requirements. Rather than
distributing computat ional workloads uniformly, the proposed framework allocated processing
resources according to prediction confidence, reserving cloud -based inference for observations
requiring additional verification. Consequently, the majority of processing decisions were completed
locally, reducing dependence on continuous network communication. Overall, the experimental
findings demonstrate the potential of confidence -guided resource allocation to improve
computational efficiency and support scalable orchard disease moni toring, particularly in
environments characterised by limited network connectivity.
4. DISCUSSION
The central methodological contribution of this study is the development of an adaptive decision-
support mechanism that determines the timing, location and execu tion pathway of inference
according to prediction confidence. This capability transforms the proposed architecture from a
conventional object detection pipeline into an intelligent decision -support system in which
computational operations are governed by the reliability of individual predictions.
The reliability of the experimental assessment was established through a dedicated annotated
orchard dataset designed to reflect representative field conditions. The collected observations
incorporated naturally occurring variations in illumination, background complexity, partial occlusion
and disease manifestation. Three alternative deployment configurations were investigated: edge -
only inference using the INT8 -quantised YOLOv8n model on the LuckFox Pico Ultra W, c loud-only
inference employing the FP32 YOLOv8l architecture, and the proposed hybrid deployment strategy.
The latter incorporated a confidence -driven decision mechanism that assessed each image before
determining its final processing pathway. Predictions associated with high confidence were accepted
directly at the edge, whereas observations exhibiting intermediate confidence were transferred to
the cloud for additional verification. Images associated with exceptionally low confidence were
rejected. This arrangement enabled computational resources to be allocated dynamically rather than
following a predetermined execution strategy. Consequently, the proposed architecture operated as
an intelligent decision -support framework rather than merely performing conv entional object
detection.
The experimental findings substantiate the effectiveness of this decision -oriented architecture.
The edge-only configuration achieved an accuracy of 75.6%, reflecting the performance limitations
associated with resource-constrained embedded hardware. Cloud-only inference improved accuracy
to 80.6%, although this improvement was accompanied by increased response latency and

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

710
substantial bandwidth requirements. In comparison, the proposed hybrid configuration achieved the
highest accuracy of 81.6%. Approximately 22.2% of the images required transmission to the cloud,
while approximately 2.2% were rejected because their prediction confidence remained below the
specified acceptance threshold. Consequently, nearly 78% of the observations were processed
locally. These results indicate that the majority of operational decisions could be completed without
requiring remote computational resources, allowing cloud infrastructure to be reserved for
observations involving greater uncertainty. Computational expenditure was therefore concentrated
on cases for which additional analysis was considered necessary.
An important characteristic of the proposed framework is its treatment of prediction confidence
as an operational decision variable rather than simply an indicator of model performance. Confidence
directly governs resource allocation, determines the location of inference and regulates
communication requirements. Through this mechanism, the system automatically selects the
appropriate computational pathway according to the reliability of its predictions. Unnecessary cloud
requests are consequently minimised, while reliable detection results remain accessible despite the
limited computational capabilities of embedded devices. In this respect, the proposed methodology
establishes a practical connection between predictive uncertainty and operational decision -making.
The decision -support capability also improves the way individual observations are handled.
Straightforward cases are resolved immediatel y through local inference, observations requiring
further examination are submitted to the cloud -based detector, and predictions with insufficient
reliability are excluded from the decision process. This hierarchical assessment strategy reduces the
likelihood of generating inappropriate decisions and reflects a reasoning process in which routine
observations receive immediate attention, whereas more challenging cases undergo additional
evaluation. Such differentiated processing cannot be readily achieved th rough conventional fixed
edge-only or cloud -only deployment, in which all observations follow the same computational
pathway irrespective of their prediction confidence.
The benefits of the proposed architecture extend beyond improvements in detection accu racy.
Compared with cloud -only deployment, the hybrid configuration reduced bandwidth consumption
by 35% and decreased end -to-end latency by a factor of 4.1. Its response time remained
comparatively close to that of edge -only execution, while its detection accuracy was comparable to
that achieved through cloud -based inference. These findings demonstrate the simultaneous
consideration of several interconnected performance objectives, including detection reliability,
response time, communication expenditure a nd computational resource utilisation. Rather than
optimising these objectives independently, the proposed decision mechanism addresses them
collectively by selecting the inference location according to prediction confidence. Local processing
is prioritised when the available confidence is sufficient, whereas cloud -based analysis is initiated
when additional computational evidence is required. Prediction uncertainty is thus translated into
actionable operational decisions.
Beyond its computational implicati ons, the proposed decision -support framework has potential
relevance to practical orchard management. Timely identification of disease symptoms can facilitate
earlier intervention and more informed crop -management decisions. Disease -related risk
information may support more targeted pesticide application and the prioritisation of inspection
activities according to the distribution of detected symptoms. Furthermore, reducing unnecessary
cloud communication can contribute to lower operational and network -related expenditure. These
characteristics make the proposed approach particularly relevant to agricultural environments where
network connectivity and computational infrastructure are limited. The principal contribution of this
investigation therefore resides in the adaptive decision -support mechanism rather than in the

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

711
implementation of hybrid computing alone. Although edge-cloud integration provides the underlying
computational infrastructure, the methodological innovation concerns the automatic determination
of when additional computational assistance is necessary and where the corresponding inference
should be executed. By linking resource allocation directly to prediction confidence, the framework
establishes a systematic relationship between detection rel iability and computational decision -
making. The same underlying principle could potentially be extended to other agricultural
applications, including pest identification, fruit quality assessment, irrigation management and
nutrient deficiency diagnosis. It s integration into autonomous agricultural platforms may also
facilitate more scalable and intelligent precision farming operations.
Nevertheless, the experimental investigation was conducted within a specific apple orchard
environment, and the collected o bservations do not encompass the full diversity of agricultural
conditions. Although the selected orchard provides a relevant setting for investigating disease
detection within Kazakhstan's apple production sector, variations in climatic conditions, illumination,
occlusion, background characteristics and disease manifestation across other agricultural regions
may influence detection performance. The proposed decision -support mechanism relies on
prediction confidence to determine the appropriate inference pa thway; therefore, its underlying
principle is potentially transferable to other crops and detection architectures. However, such
applications require the corresponding crop -specific models to undergo appropriate training and
validation before deployment. S imilarly, adaptation to alternative embedded platforms requires
device-specific calibration to accommodate differences in computational capacity and
communication constraints. The confidence thresholds may also require recalibration to reflect
different en vironmental and operational conditions. Accordingly, although the proposed
methodology offers potential for broader edge -cloud deployment, its reliability across diverse
agricultural environments must be established through further validation. Future resea rch will
therefore focus on expanding dataset coverage and systematically evaluating cross -domain
generalisation and deployment across different agricultural settings
5. CONCLUSION
This study developed an adaptive decision-support system for automated apple disease detection
using a hybrid edge -cloud architecture. A confidence -based inference mechanism enabled high -
confidence predictions to be processed locally, uncertain cases to be transferred to the cloud, and
low-confidence predictions to be rejected. An INT8-quantised YOLOv8n model deployed on the low-
cost LuckFox Pico Ultra W achieved over 20 FPS with an average edge latency of 231 ms. Cloud
processing employed YOLOv8l through Wi-Fi, HTTP and FastAPI, with only 22.2% of images requiring
additional analysis. The system achieved 81.6% detection accuracy, a 6% improvement over edge -
only inference, while reducing bandwidth consumption by 35% and end-to-end latency by a factor of
4.1 compared with cloud -only processing. The principal contribution lies in transl ating prediction
confidence into intelligent computational decisions, enabling selective cloud utilisation and efficient
resource allocation. This approach offers a scalable solution for orchard monitoring under limited
connectivity and computational resou rces, with broader applications in precision agriculture and
Internet of Things environments. Future research should explore learned decision policies, adaptive
confidence thresholds and multimodal sensing incorporating RGB, thermal, humidity and soil data .
Federated learning and fog computing could enhance distributed intelligence, while neural
architecture search and Tiny Vision Transformers may improve embedded inference efficiency.
Cloud-to-edge knowledge transfer and continuous model updating could fur ther strengthen
detection reliability under changing network conditions, disease prevalence and orchard
environments.

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

712
FUNDING
This work was supported by the Science Committee of the Ministry of Science and Higher
Education of the Republic of Kazakhstan (Gr ant No. AP23488699, "System for Detecting
Phytopathogenic Infections in Domestic Fruit Crop Production Using Computer Vision and Artificial
Olfaction").

CONFLICTS OF INTEREST
The authors declare no conflicts of interest.

References
[1] Savvidis, P ., & Papakostas, G. A. (2021). Remote crop sensing with iot and AI on the edge In
2021 IEEE World AI IoT Congress (AIIoT), https://doi.org/10.1109/aiiot52608.2021.9454237
[2] da Silva, J. C. F., Silva, M. C., Luz, E. J. S., Delabrida, S., & Oliveira, R. A. R. (2023). Using mobile
edge AI to detect and map diseases in citrus orchards. Sensors, 23(4), 2165.
https://doi.org/10.3390/s23042165
[3] Yuchao, X., Qian, W., Bingyuan, Z., Lingli, Z., Ni, R., & Meina, Z. (2025). Review on lightweight
deep learning networks for object detection in crops. Journal of Chinese Agricultural
Mechanization, 46(3), 261. https://doi.org/10.13733/j.jcam.issn.2095-5553.2025.03.038
[4] Grujev, M., Stojanovic, D., Milosavljevic, A., Ilic, M., & Prodanovic, V. (2026). An efficient real-
time apple disease classification approach usi ng a novel lightweight neural network on an
arduino edge device. Internet of Things , 35, 101833.
https://doi.org/10.1016/j.iot.2025.101833
[5] Zheng, Z., Zhang, Y ., & Sun, L. (2025). Improved yolov8-based algorithm for citrus leaf disease
detection. IEEE Access, 13, 105888-105900. https://doi.org/10.1109/access.2025.3579074
[6] Zhou, S., Yin, W., He, Y ., Kan, X., & Li, X. (2025). Detection of apple leaf gray spot disease based
on improved yolov8 network. Mathematics, 13(5), 840.
https://doi.org/10.3390/math13050840
[7] Liu, B., Fan, H., Zhang, Y ., Cai, J., & Cheng, H. (2024). Deep learning architectures for diagnosing
the severity of apple frog -eye leaf spot disease in complex backgrounds. Frontiers in Plant
Science, 14. https://doi.org/10.3389/fpls.2023.1289497
[8] Wang, G., Sang, W., Xu, F., Gao, Y ., Han, Y ., & Liu, Q. (2025). An enhanced lightweight model
for apple leaf disease detection in complex orchard environments. Frontiers in Plant Science,
16. https://doi.org/10.3389/fpls.2025.1545875
[9] Zou, H., Lv, P ., & Zhao, M. (2025). Detection of apple leaf diseases based on lightYOLO -
AppleLeafDx. Plants, 14(4), 599. https://doi.org/10.3390/plants14040599
[10] Liu, Y ., Zhang, Y ., Jiang, D., Zhang, Z., & Chang, Q. (2023). Quantitative assessment of apple
mosaic disease severity based on hyperspectral images and chlorophyll content. Remote
Sensing, 15(8), 2202. https://doi.org/10.3390/rs15082202
[11] Karim, M. J., Goni, M. O. F., Nahiduzzaman, M., Ahsan, M., Haider, J., & Kowalski, M. (2024).
Enhancing agriculture through real -time grape leaf disease class ification via an edge device
with a lightweight CNN architecture and Grad -CAM. Scientific Reports , 14(1).
https://doi.org/10.1038/s41598-024-66989-9
[12] Manoj, R. K., & Nandhini, S. A. (2025). C omputer vision and AI -based apple plant disease
detection framework in smart farming. In Studies in Computational Intelligence (pp. 269-286).
Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-05603-0_12
[13] Ramya, R. K., & Meenakshi, C. (2025). An adaptive multi -stage ensemble learning approach

Decision Making: Applications in Management and Engineering
V olume 9, Issue 1 (2026) 685-713

713
for accurate diagnosis of tomato leaf diseases In 2025 6th International Conference on
Inventive Research in Compu ting Applications (ICIRCA),
https://doi.org/10.1109/icirca65293.2025.11089636
[14] Rehman, S., Attique Khan, M., Alhaisoni, M., Armghan, A., Alenezi, F., Alqahtani, A., Vesal, K.,
& Nam, Y. (2023). Fruit leaf diseases classification: A hierarchical deep learning framework.
Computers, Materials & Continua , 75(1), 1179 -1194.
https://doi.org/10.32604/cmc.2023.035324
[15] Kannan, E., M J, C. M. B., S, A. D., N, R. N., Begum, A., & D, H. (2024). Deep learning techniques
advancements in apple leaf disease detection. Procedia Computer Science , 235, 713 -722.
https://doi.org/10.1016/j.procs.2024.04.068
[16] Dhiman, P ., Kaur, A., Hamid, Y ., Alabdulkreem, E., Elmannai, H., & Ababneh, N. (2023). Smart
disease detection system for citrus fruits using deep learning with edge computing.
Sustainability, 15(5), 4576. https://doi.org/10.3390/su15054576
[17] Dhiman, P ., Manoharan, P ., Lilhore, U. K., Alroobaea, R., Kaur, A., Iwendi, C., Alsafyani, M.,
Baqasah, A. M., & Raahemifar, K. (2023). PFDI: A precise fruit disease identification model
based on context data fusion with faster -CNN in edge computing environment. EURASIP
Journal on Advances in Signal Processing, 2023(1), 1-18. https://doi.org/10.1186/s13634-023-
01025-y
[18] Foroughi, A., Lloret, J., Jimenez, J. M., & Sendra, S. (2025). An edge computing wireless sensor
network for diagnosing orange fruit disease. Cluster Computing , 28(5).
https://doi.org/10.1007/s10586-024-04999-y
[19] Thongnim, P ., Sreekajon, J., & Pukseng, T. (2025). Optimizing fruit cultivation through an
unsupervised learning approach with data scaling techniques for strategic insights. Eurasian
Journal of Mathematical and Compute r A

论文原文与出版信息

Decision Making Applications in Management and Engineering
查看论文原文

继续探索