Enhanced daily CHIRPS precipitation using sequential K-nearest neighbors correction and Kalman filter blending for Thailand

Accurate precipitation data is essential for hydrological modeling and water resource management, particularly in tropical regions with complex topography and limited ground-based observation networks. This study develops an integrated two-stage framework combining K-nearest neighbors (KNN) machine learning bias correction with Kalman filter blending to enhance Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) daily precipitation estimates across Thailand's diverse geographical and climatic conditions. The methodology utilized comprehensive meteorological data from 628 stations across Thailand spanning 44 years (1981-2024), with temporal partitioning into training (1981-2015) and validation (2016-2024) periods. The first stage implemented seasonal KNN bias correction using 11-dimensional feature vectors incorporating CHIRPS satellite precipitation, auxiliary meteorological variables (maximum/minimum temperature, relative humidity, evaporation), and station coordinates. The second stage applied adaptive Kalman filter blending with dual-update processing, combining raw CHIRPS data with KNN-corrected estimates. Results demonstrate exceptional performance improvements across both periods. Correlation coefficients increased dramatically from 0.42 to 0.94 during training (124% improvement) and from 0.41 to 0.91 during validation (122% improvement). Systematic bias correction transformed raw CHIRPS overestimation of 34.03% to controlled underestimation of -10.00% (BC CHIRPS) and -10.78% (BBC CHIRPS) during training, with similar validation patterns. Regional analysis revealed differential effectiveness across Thailand's climatic zones. The most challenging DJF dry season showed severe raw CHIRPS overestimation of 588.27% (training) and 167.12% (validation), reduced by 95-99% with both corrections. Spatial validation confirmed operational applicability, effectively eliminating widespread overestimation while preserving legitimate precipitation signals. The integrated framework successfully addresses systematic biases in satellite precipitation products while maintaining computational efficiency. This research demonstrates that sophisticated machine learning integrated with optimal filtering theory can significantly enhance satellite precipitation accuracy for operational applications in data-scarce tropical regions, with demonstrated effectiveness across Thailand's diverse conditions and strong potential for broader tropical applications.