Exploring Relationships Between Drought Indices and Ecological Drought Impacts Using Machine Learning and Explainable AI

Abstract

Rangeland ecosystems in the United States hold great ecological, economic, and cultural value. However, the increasing frequency and severity of droughts pose potential threats to these ecosystems. This study used remotely sensed data, machine learning, and explainable artificial intelligence (XAI) to explore relationships between drought indices and vegetation health in the Cheyenne River Basin, USA. The study employed XGBoost Regressor and Extra Trees Regressor models in conjunction with SHapley Additive exPlanations (SHAP) to identify the most influential drought indices and environmental variables for predicting Normalized Difference Vegetation Index (NDVI), thereby uncovering indicators of vegetation stress in the basin. The XGBoost regressor was moderately successful at predicting NDVI, making the model suitable for subsequent XAI analysis using SHAP. SHAP results revealed that the Palmer Drought Severity Index (PDSI), the 90-day Standardized Precipitation Index (SPI), and snow water equivalent (SWE) were the most influential predictors of NDVI, indicating their strong association with changes in vegetation health in the Cheyenne River Basin. This study demonstrates the feasibility and value of applying XAI methods to investigate both the strength and directionality of ecological drought indicators—an approach that has been underutilized in drought research. These insights can inform future drought research, improve monitoring efforts, and help anticipate ecological drought impacts in the region.