A Lightweight ConvNeXt Framework for Cotton Plant Disease Detection and Pest Identification Toward Sustainable Precision Agriculture
Abstract
Cotton (Gossypium spp.) ranks among the most significant cash and fibre crops globally; nonetheless, its productivity is persistently compromised by foliar diseases and insect pest infestations that, if not promptly addressed, can lead to yield reductions of 20–50%. Conventional field scouting is labour-intensive, subjective and often too slow to support targeted intervention. This research introduces an automated, image-centric deep-learning architecture that concurrently tackles two interrelated tasks: (i) multi-class cotton leaf disease detection across seven categories and (ii) cotton pest identification across five categories, using two publicly available field-image datasets. A lightweight ConvNeXt-Tiny architecture is introduced, featuring a frozen ImageNet-pretrained backbone paired with a streamlined custom classification head that includes global average pooling, layer normalisation, dropout, and a softmax layer, allowing for the training of merely a few thousand parameters. The proposed model is benchmarked against five widely used baselines — a custom CNN, an LSTM, InceptionV3, ResNet50 and a Vision Transformer (ViT) — under an identical preprocessing, augmentation and 80/20 validation protocol. On the disease dataset the proposed Lightweight ConvNeXt achieved 95% accuracy, precision, recall and F1-score, outperforming CNN (93%), InceptionV3 (89%), ViT (86%), and substantially exceeding ResNet50 (58%) and LSTM (59%). On the pest dataset it achieved 98% across all metrics, ahead of InceptionV3 (95%), CNN (94%), ViT (92%), ResNet50 (88%) and LSTM (84%). Critically, these gains were obtained while training only 6,919 parameters, several orders of magnitude fewer than the fully trained CNN. The results demonstrate that transfer learning with a frozen modern-convolutional backbone offers an accurate, data-efficient and deployment-friendly solution for cotton crop health monitoring and can serve as a decision-support component within precision-agriculture and integrated pest-management systems.