CAUSAL FEATURE ATTRIBUTION AND UNCERTAINTY CALIBRATION IN SPATIOTEMPORAL MACHINE LEARNING SYSTEMS

Spatiotemporal machine learning systems increasingly inform forecasts, risk assessments, and operational interventions in settings where observations are interdependent, measurement conditions vary, and errors have unequal consequences across locations. Interpretability and uncertainty are frequently addressed through separate technical routines, although an attribution can be misleading when input variables reflect common causes, while a calibrated prediction may conceal unstable dependence on environmental or institutional conditions. This article develops a methodological framework for jointly examining causal feature attribution and uncertainty calibration in spatially and temporally structured learning. Its central distinction separates predictive influence, evidence of a causal mechanism, and warranted confidence in a forecast. The framework specifies how to define a target decision, document plausible causal relations, construct spatial and temporal validation splits, compare conditional and intervention-oriented explanations, and assess calibration at operationally meaningful levels of aggregation. Attention is given to spatial interference, delayed labels, uneven sensor coverage, extreme events, and the ethical implications of communicating geographically differentiated confidence. Three qualitative tables organize identification assumptions, a validation matrix, and reporting requirements. The analysis synthesizes established causal reasoning, dependence-aware explanation methods, calibration research, proper scoring principles, and spatiotemporal Earth-system modelling. No new empirical experiment is claimed. A reliable workflow is characterized by transparent assumptions, diagnostic contrasts, explicit limits on causal identification, and governance procedures that connect model explanations to the decisions they are intended to support.