Enhanced Weed Detection in Agricultural Fields Using Convolutional Neural Networks and SHAP Interpretability Techniques

Abstract

Effective weed detection is crucial for optimizing agricultural productivity and sustainability. This study proposes an innovative approach to enhance weed detection in agricultural fields using customized Convolutional Neural Networks (CNNs) and SHAP (SHapley Additive exPlanations) interpretability techniques. Leveraging recent advancements in deep learning and model interpretability, our method integrates a customized CNN model with SHAP to achieve high accuracy in weed detection and providing optimal results. Though the data set is unbalanced, over four classes but the minority classes are given adjustment weights with the model performance has improved to 0.75 accuracy.  And this study also percents the modified VGG model with adjusted weights, and achieved an accuracy of 0.98. The results are interpreted with SHAP, this enables effectiveness of the approach. For this approach DeepWeeds dataset is used and tested.