End-to-End Jute-Pest Detection By Explainable Lightweight CNN

Owing to their severity and widespread incidence, which result in large crop losses, jute pests are regarded as the main issue for agricultural output. Soft computing techniques can be used to protect crops from hazardous pests, which is important for increasing agricultural yield. The deep learning (DL) and conventional machine learning (ML) methodologies serve as the foundation for soft computing solutions. However, DL approaches are computationally expensive and require a huge quantity of training data. In contrast, the selection of manual feature extraction processes in standard methods is inefficient, time-consuming, and unsuccessful. This research presents an effective approach for detecting pests using a jute that accurately classifies the bugs and visualizes them based on the specified class label by Shapley Additive Explanations (SHAP). In the proposed work, we proposed a Lightweight CNN model. Experimental results demonstrate that it achieved 99.43 % accuracy outperforming existing state-of-the-art (SOTA) models. The paper further focuses on the interpretability aspect by employing SHAP visualizations, shedding light on the key features influencing model predictions and thus enhancing transparency in decision-making. The end-to-end integration with Gradio not only amplifies the model's utility but also facilitates broader accessibility, allowing stakeholders to engage with the technology in real time. We compare the suggested model's results with those of numerous SOTA models to assess the model's efficacy; the suggested model produced the best outcomes in the experiments. We created the suggested model end-to-end using Gradio since it has the potential to be used in practical settings. This will encourage more studies on pest identification to boost agricultural output.