A Comparative Study of Transfer Learning Models for Cotton Plant Anomaly Identification with Explainable AI

Abstract

The health of the cotton crop is at risk due to numerous ailments, such as diseases and pests, which are conventionally difficult and prone to human error in diagnosis by hand. In this paper, we have provided a comparative analysis of four popular transfer learning frameworks: ResNet50, MobileNetV2, DenseNet121, and EfficientNetB3 to automatically identify six main cotton plant anomalies. Evaluation and training were done on 4200 images, which were divided into three diseases (Bacterial Blight, Powdery Mildew, Target Spot), two pest problems (Aphids, Army Worm), and healthy leaves. The models were properly evaluated in terms of accuracy, precision, recall, and F1-score. The findings indicate that EfficientNetB3 is state-of-the-art with an accuracy of 99.58% and DenseNet121 comes next with 98.75%. To overcome the black box character of these models, we have used two Explainable AI (XAI) methods, LIME and Grad-CAM, to understand the reason behind their predictions. The XAI visualizations also validated the decision-making process of the models based on the features that were considered by them as relevant in relation to pathology and pest related features. This research shows that EfficientNetB3 outperforms the other two systems and presents the importance of XAI to make strong, reliable, and deployable diagnostic systems to use in precision farming.