Comparative Evaluation of XAI Methods for Transparent Crop Yield Estimation Using CNN

While Deep Learning (DL) has significantly improved the estimation of crop yields from satellite imagery, the intricacies of the decision-making processes within the utilized DL models remain obscure. To ensure transparency and dependability in crop yield predictions, unravelling the complex mechanisms of DL models is imperative. This task is challenging due to the varying outcomes presented by different eXplainable Artificial Intelligence (XAI) methods. This study examines a range of eXplainable Artificial Intelligence (XAI) techniques applied to Convolutional Neural Networks (CNNs) for Soybean yield estimation from Sentinel-2 satellite imagery. Methods like Layerwise Relevance Propagation (LRP), SmoothGrad, Deep Taylor, and gradCAM have been employed to elucidate the CNN model, producing saliency maps that are subsequently evaluated through a perturbation analysis. The study also examines how each XAI method identifies the model's focus on Soybean fields compared to other crop fields or land uses. Our findings indicate that LRP outperforms other methods, offering more accurate saliency maps and highlighting critical spatial information for crop yield estimation. The knowledge gained about the XAI methods facilitates understanding the behaviour of complex CNN model architecture used for crop yield estimation in the future.