SHAP-guided Explanations for the Machine Learning Classification of Irrigated Fields Using Satellite Imagery

Abstract

ntegration of machine learning (ML) models into real-world applications such as agriculture, demands explanations of the inner workings of such models. As ML models evolve and become more sophisticated, it becomes more challenging for humans to comprehend their decision-making processes. This lack of transparency presents a significant barrier to the widespread adoption of ML-driven models, especially in precision agriculture where the stakes are high. One of the applications in precision agriculture is the creation of ML-driven models for detecting irrigated fields. To provide explanations for the ML classifiers for detecting irrigated fields based on satellite imagery, we employ explainable AI (xAI), a paradigm that seeks to provide comprehension of the decision-making process hidden behind predictions given by the ML models. By using the SHAP method, we seek to unravel how the model decisions differ among several crops of interest (maize, soybean, and sugar beet) for the classification task of detecting irrigated fields. On the more fine-grained level, we seek to analyse the most decisive factors among fields covered with the same crop. Results indicate that the main factors (vegetation indices for different days during the season) that guide the decision of ML models differ among different crops. More importantly, the results show that some parts of the decision-making process of the ML models deviate from human reasoning. More sensible classification of irrigated fields is achieved for sugar beet, than in cases for soybean and maize. Analysis of the explanations why particular fields are classified as irrigated or not, confirmed conclusions obtained on the crop-specific explanations. Our results show that having syntheses of xAI and ML classification models for irrigated fields is a crucial component for building trust, mitigating biases, and enhancing the robustness of these models.