Contrastive Self-Supervised Learning for Cocoa Disease Classification

Abstract

Self-supervised learning has shown significant improvement within several disciplines including the health sectors and areas where acquisition of annotated data is difficult. This study focuses on the application of self-supervised learning in tackling cocoa disease detection and classification challenges by taking cocoa swollen shoot virus disease (CSSVD) and anthracnose. Self-supervised learning has proven to offer resilience against both OOD and adversarial attacks and as an alternative to the cost involved in the annotation process of the vast amount of data. This work leverages a contrastive self-supervised learning pretext model trained on cocoa leaves and pods, a remarkable performance was shown when we fine-tuned and pre-trained on a downstream task. Several works are being done on the subject of self-supervised learning and fully-supervised learning. In order to get the best out of these two techniques, we propose a hybrid approach which will combine both SLL and fully supervised learning. Thus in any deep learning project regardless of the amount of annotated dataset around, a pretext model should be developed from the dataset which can then be fined-tuned using the entire dataset at hand. Three models were developed and the best come in as a result of fine-tuning the learnt features using the entire dataset. After performing the downstream task (classification of cocoa diseases) a validation accuracy of 87%, 72% and 69% for the fine-tune, fully-supervised and pre-trained respectively were obtained. This approach holds promise for enhancing the productivity in crop management, particularly in regions like Ghana and Ivory Coast where their GDP relies on cocoa production © 2024 IEEE.