Standalone SqueezeNet for Accurate and Resource-Efficient Tomato Leaf Disease Classification

Abstract

In agricultural productivity, there is a decrease in crop yield due to tomato leaf disease. An early detection system with high accuracy is needed to address this issue. This study evaluates the performance of a standalone SqueezeNet-based Convolutional Neural Network (CNN) architecture for identifying five tomato leaf types using the PlantVillage dataset. In contrast to approaches in the literature that often use complex models with large image features, this study demonstrates that an architecture with smaller feature extraction can achieve a stable level of precision through data processing optimization and K-fold cross-validation. The methodology includes 96x96 pixels input resolution, bilinear interpolation, and data normalization with 1/255 scale augmentation to strengthen the model’s generalization. Model reliability was tested using the 5-fold stratified cross-validation method. Experimental results show that the standalone SqueezeNet successfully achieved high classification accuracy of 97.97%, which slightly increased to 98.02% after 8-bit quantization. The final model occupies only 0.78 MB of Flash memory with an average inference latency of 2.37 ms per image. In-depth analysis through a confusion matrix demonstrated the model’s consistent ability to distinguish visually similar disease symptoms, such as Late Blight and Leaf Mold. This approach demonstrates that SqueezeNet can effectively extract features from images of plant diseases, leading to efficient classification and detection of tomato diseases. The effectiveness of this plant disease classification model could establish a new standard approach that can be used without requiring complicated model architectures.