CA-ASPP-ConvNeXt: A Lightweight Multi-Scale Framework for Tomato Leaf Disease Classification Model

Abstract

Early and accurate detection of tomato leaf diseases is essential for sustainable agriculture and crop yield improvement. In this study, we propose a novel deep learning-based classification model that combines the strengths of transfer learning, attention mechanisms, and multi-scale feature extraction. The backbone of the model utilizes ConvNeXt-Tiny, pretrained using self-supervised learning techniques such as SimCLR or BYOL, to extract rich hierarchical features from input tomato leaf images of size 224×224×3. To enhance the representational capacity of the network, Coordinate Attention (CA) modules are integrated after each stage of the ConvNeXt backbone, allowing the model to focus on both spatial and channel-wise informative features. To capture multi-scale contextual information critical for identifying fine-grained disease patterns, the network is further augmented with a Atrous Spatial Pyramid Pooling (ASPP) module. The resultant feature map is passed through a Global Average Pooling (GAP) layer, followed by a Dropout regularization and a fully connected Softmax classifier to perform multi-class classification across 11 distinct tomato leaf disease categories. Extensive experiments on a tomato leaf disease dataset demonstrate that our model achieves high classification accuracy while maintaining computational efficiency, making it suitable for real-time agricultural diagnostics and mobile deployment. The integration of coordinate attention and self-supervised pretrained

features significantly boosts performance, particularly in distinguishing visually similar diseases under varied environmental conditions.