Lightweight squeeze U-net for pixel-wise classification of cassava anthracnose disease on edge device

This paper introduces a lightweight deep learning model, a modified Squeeze U-Net for pixel-level segmentation of Cassava Anthracnose Disease on edge devices. Cassava, a critical crop for Nigeria, suffers significant yield losses due to anthracnose, which causes leaf distortions and spots. Traditional visual inspection methods are inefficient and error-prone, emphasizing the need for automated disease detection. Building on the architectures of U-Net and Squeeze U-Net, our approach incorporates innovative modifications that drastically reduce model complexity. The Modified Squeeze U-Net achieves a 48-fold reduction in size (7.11 MB compared to 386 MB for U-Net), and a significant decrease in parameter count while maintaining competitive accuracy (approximately 92–98%). Evaluations were performed using a dataset of 500 annotated cassava leaf images, with testing on a Raspberry Pi 3 demonstrating real-time, on-field detection without reliance on cloud infrastructure. Comparative analysis reveals that while U-Net provides the highest accuracy, its large size renders it impractical for resource-constrained environments, and Squeeze U-Net’s efficiency comes at the cost of lower accuracy. In contrast, the Modified Squeeze U-Net strikes an optimal balance between performance and efficiency, making it particularly well-suited for precision agriculture applications. Future work will assess the model’s effectiveness across different crops and its integration into portable devices such as smartphones.