An Explanation Technique For Yield Prediction in Smart Farming

Sandya De Alwis, Bahadorreza Ofoghi, Yishuo Zhang

Abstract: The utilization of artificial intelligence tools and methods in agriculture has increased over the last few years. However, farmers and agronomists are uncertain about trusting such tools and the findings of machine/deep learning model predictions. This work develops a novel eXplainable AI (XAI) technique for smart farming yield prediction that can more effectively explain prediction outcomes when compared with state-of-the-art current explainers, LIME and SHAP. We call the proposed model, which considers both attributes and time lag, the Duo Attention eXplainable Mechanism (DAXM). We have developed and tested the model with two separate farming data sets and the results of our experiments demonstrate the effectiveness of prediction features for a three-week window on both tomato and strawberry yield prediction. We show that the explanation of such features can be achieved more effectively through our proposed DAXM model while these explanations significantly differ at the 95% confidence level from those generated by LIME and SHAP. DAXM is also more aligned with expert opinion with a higher degree of agreement with expert-reported feature importance measures as compared with LIME and SHAP. The proposed XAI approach for smart farming yield prediction offers effective explanations that can enrich user adaption of the cutting-edge neural models in the domain.