Towards interpretability: Assessment of residual networks for tomato leaf disease classification

The tomato occupies a prominent place in the Philippines' agricultural economy. However, tomato leaf diseases are challenges in tomato crop production leading to economic losses. Among the tomato leaf diseases, early blight and Septoria leaf spot are prevalent in the Philippines due to the climate. Thus, the accurate identification of diseases affecting tomato leaves is essential. Currently, a visual inspection is the primary method for diagnosing tomato leaf diseases which is time-consuming and inefficient. This study aims to develop a quantized Residual Network with convolutional 50 layer (ResNet-50) based model to classify tomato leaves as healthy or affected by Septoria leaf spot or early blight. Furthermore, to enhance the reliability of the models' classification, gradient-weighted class activation mapping (Grad-CAM) was implemented. In contrast with the visual inspection, a programmed system does not get tired and can provide consistent performance results. As a result, the original 32-bit floating point model attained an accuracy rate of 91.22%. The quantized 16-bit floating point model demonstrated comparable performance with 90.10% accuracy with a 50% reduction in the model size and inference time of 0.3942 seconds. The minimal accuracy loss of the 16-bit model relative to the 32-bit model is due to the post-training quantization. The reduction to 16-bit precision is significant for the future deployment of edge devices where resources are limited.