Spatiotemporal transformer modeling of satellite fire detection confidence under climate variability

This study investigates multi-class modeling of VIIRS satellite fire detection confidence levels (low, medium, high) using twelve years (2012-2023) of thermal, radiative, spatial, and temporal observations over Saudi Arabia. Detection confidence is formulated as a reliability-oriented structured learning problem. Transformer architectures with distinct representation and reasoning paradigms are systematically compared against conventional machine-learning baselines under a temporally stratified validation protocol to ensure robustness across interannual variability. Transformer-based models demonstrate substantially improved discrimination and regression-consistency performance relative to classical baselines. Qwen3 achieves the highest overall accuracy (0.999), F1-score (0.993), and [Formula: see text] (0.988), with minimal prediction error (RMSE = 0.032; MAE = 0.001) and strong probabilistic calibration (ECE 0.0008; Brier 0.0005). In contrast, ensemble baselines achieve accuracies between 0.956 and 0.964, indicating that while radiative attributes provide strong predictive signals, contextual self-attention mechanisms significantly enhance structured confidence differentiation. Longitudinal analysis reveals stable performance across seasonal and interannual climatic variability, suggesting that detection confidence exhibits consistent nonlinear dependencies within radiative-spatiotemporal feature space. The consistent performance across varying climatic conditions further suggests that the model captures invariant spatiotemporal and radiative structures rather than region-specific patterns, supporting its potential for cross-regional generalization. Feature attribution analysis further confirms that thermal radiative measurements and acquisition timing are dominant contributors to confidence stratification. These findings provide empirical evidence that foundation transformer architectures effectively model structured detection reliability in remote sensing systems, offering a scalable framework for probabilistic and uncertainty-aware environmental intelligence under climate variability.