CNN-Based Hybrid Model for Detecting Blight Diseases in Potato Crops with Advanced Image Processing Techniques

Abstract

Potato production plays a vital role in global agriculture as a major food source for large populations. However, potato crops are highly susceptible to diseases, particularly Early Blight and Late Blight, which result in substantial yield losses. Timely detection and effective control of these diseases are essential for maintaining stable crop output. This study explores the integration of Convolutional Neural Networks (CNNs) and advanced image processing techniques to differentiate between diseased and healthy potato plants accurately. Two datasets comprising original and enhanced images were used to train four CNN models: InceptionV3, Xception, Densenet201, and Resnet152V2. The original images underwent background removal only, whereas the enhanced images were further processed using contrast enhancement and morphological transformation in addition to background removal to reduce noise, improve quality, and prepare the images for analysis. The CNN models were trained using these datasets, with their bottom layers fixed and the top layers fine-tuned to improve performance and reduce training time. Experimental results revealed that models trained on enhanced images achieved a 2.45% to 4.45% improvement in accuracy, precision, and sensitivity compared to those trained on original images. Moreover, a hybrid model that combined two high- performing CNNs achieved a 98.91% accuracy, marking up to 10.69% improvement over individual models. This approach offers significant potential for reducing crop yield losses while minimizing dependence on chemical treatments.