Advancements in Potato Leaf Disease Detection Using a Deep Convolutional Neural Network with Enhanced CLAHE-Based Preprocessing and Channel Attention

Abstract

Potato (Solanum tuberosum) is a globally important food crop whose yield losses, caused by foliar diseases including early blight, late blight, and a range of bacterial, fungal, viral, nematode, and pest-induced conditions, are often difficult to identify in the field and costly to diagnose. This paper introduces a potato-leaf disease detection framework based on a Deep Convolutional Neural Network (DCNN) that integrates a lightweight depthwise-separable convolutional backbone, a squeeze-and-excitation channel-attention module, and a carefully designed preprocessing pipeline. The images are preprocessed using quality-enhancement techniques such as Contrast-Limited Adaptive Histogram Equalization (CLAHE) to reduce background noise, an edge-preserving denoising algorithm to enhance the texture of disease-relevant regions of the leaf, and colour normalization together with leaf segmentation to remove background clutter before the images are fed into the network. The framework is described and evaluated on two open-access, validated datasets: the potato subset of PlantVillage, comprising 2,152 images across three classes acquired under controlled conditions, and the dataset of Shabrina et al., comprising 3,076 images across seven classes acquired from real agricultural fields under uncontrolled conditions. The proposed attention-augmented DCNN with enhanced preprocessing is compact enough for edge deployment and achieves better classification performance than baseline CNN, VGG-16, ResNet-50, MobileNetV3, and EfficientNet-B0 models in terms of accuracy, macro-averaged F1-score, and area under the ROC curve.