Attentive Self-supervised Contrastive Learning (ASCL) for plant disease classification

Deep-learning plays a crucial role in large-scale health monitoring of agricultural plants. One of the challenges in plant disease classification is the limited availability of annotated training data, where supervised deep feature learning typically excels. However, traditional deep learning backbones often produce representations that are not sufficiently discriminative or interpretable for detailed plant disease analysis. We propose an Attentive Self- supervised Contrastive Learning (ASCL) framework that leverages transferable representations as supervision signals. The ASCL framework enhances interpretability by incorporating attention mechanisms, such as squeeze- excitation and convolutional block attention module, which highlight key regions in plant images, aiding in transparent decision-making. In the present work, a pre-trained squeeze-excitation ResNet50 Siamese backbone network on the unlabeled PlantVillage dataset was used to validate the generalizability of the learned representations. The pre-trained weights were then fine-tuned on small-scale unseen datasets extracted from the 17- class PlantVillage Taiwan Tomato and Apple datasets. Despite involving fewer than 17 classes, the high variability within each class, such as the disease progression stages, underscores the fine-grained nature of the classification task. Extensive experiments demonstrated that the ASCL framework achieved 89 % accuracy, outperforming a baseline supervised model that scored 88.9 %. Moreover, when the ASCL learned, weights were transferred to an unseen dataset, and the model achieved 93.5 % accuracy, compared to 91.66 % with supervised ResNet50. The framework is scalable to larger datasets with more classes, making it applicable to broader finegrained classification tasks. Therefore, the proposed ASCL framework demonstrates the generalizability and transferability of the downstream plant disease classification tasks.