<p>Accurate identification of plant leaf diseases is critical for precision agriculture, as it ensures healthy crop yields and reduces losses. Although numerous state-of-the-art deep learning models based on different architectures, such as convolutional neural networks, Transformers, and graph neural networks, have shown outstanding performance in general classification tasks, their effectiveness in specific applications, such as plant disease classification and recognition, has not been extensively explored. Therefore, evaluating these models is essential to identify the most effective approaches for the task of leaf disease classification and recognition, particularly by leveraging the advantages of pretraining on large-scale datasets such as ImageNet. Moreover, lightweight models are needed for deployment in precision agriculture, as they enable real-time processing on edge devices, ensuring timely interventions in the field. This study evaluates multiple fine-tuned deep learning models based on the three aforementioned architectures, achieving accuracy from 89.30 to 98.70% on several public leaf disease datasets. Additionally, we created a new cucumber leaf dataset with four classes, including a healthy leaf category, comprising a total of 8057 samples, and evaluated models’ performance on this dataset as well. To optimize deployment, we applied post-training quantization to the fine-tuned models, observing only a slight decrease in performance of Transformer-based models from 0.49 to 1.62% while achieving <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="521_2025_11400_Article_IEq1.gif" Format="GIF" Height="6" Rendition="HTML" Resolution="72" Type="Linedraw" Width="17" /> </InlineMediaObject> <EquationSource Format="TEX">$\tiny {\sim }$</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>  <b>4</b><InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="521_2025_11400_Article_IEq2.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> reduction in model storage requirements. The findings suggest that these lightweight models, with their balanced trade-off between accuracy and storage, can be adopted for real-time deployment on edge devices in precision agriculture.</p>

Compact deep learning models for leaf disease classification and recognition in precision agriculture

期刊论文
发表日期：

2025年6月16日

查看全文

Ishwar Chandra Mahto,
Jimson Mathew

摘要

Accurate identification of plant leaf diseases is critical for precision agriculture, as it ensures healthy crop yields and reduces losses. Although numerous state-of-the-art deep learning models based on different architectures, such as convolutional neural networks, Transformers, and graph neural networks, have shown outstanding performance in general classification tasks, their effectiveness in specific applications, such as plant disease classification and recognition, has not been extensively explored. Therefore, evaluating these models is essential to identify the most effective approaches for the task of leaf disease classification and recognition, particularly by leveraging the advantages of pretraining on large-scale datasets such as ImageNet. Moreover, lightweight models are needed for deployment in precision agriculture, as they enable real-time processing on edge devices, ensuring timely interventions in the field. This study evaluates multiple fine-tuned deep learning models based on the three aforementioned architectures, achieving accuracy from 89.30 to 98.70% on several public leaf disease datasets. Additionally, we created a new cucumber leaf dataset with four classes, including a healthy leaf category, comprising a total of 8057 samples, and evaluated models’ performance on this dataset as well. To optimize deployment, we applied post-training quantization to the fine-tuned models, observing only a slight decrease in performance of Transformer-based models from 0.49 to 1.62% while achieving $\tiny {\sim }$ $\sim$ 4 $\times $ $\times$ reduction in model storage requirements. The findings suggest that these lightweight models, with their balanced trade-off between accuracy and storage, can be adopted for real-time deployment on edge devices in precision agriculture.