Integrating remote sensing and random forest for accurate prediction of soil and biomass phosphorus dynamics in rice fields across complex terrain

Phosphorus detection remains challenging despite advanced methods, especially with complex environmental factors across varied terrains. Phosphorus detection highlights the need to enhance soil security to reduce fertilizer overuse and policies. This study combines Random Forest analysis with remote sensing to detect soil available phosphorus (SAP), total phosphorus biomass (TPB), and phosphorus uptake efficiency (PUE). The study was conducted in Malang Regency, East Java, Indonesia, where point observations were taken in volcanic, alluvial, and karst terrains. All three phosphorus indicators, SAP, TPB, and PUE, were analyzed using Random Forest models that incorporated a comprehensive set of environmental covariates, including topographic attributes, soil properties, climatic variables, and vegetation indices derived from remote sensing.. Performance optimization was done through hyperparameter tuning, with accuracy assessed via R², RMSE and RPIQ. The models demonstrated strong performance, with R² values of 0.928 for SAP, 0.927 for TPB, and 0.922 for PUE. The corresponding RMSE values were 10.192, 5.197, and 27.813, respectively. RPIQ scores of 1.19 (SAP), 2.45 (TPB), and 1.43 (PUE) further indicate reliable predictive accuracy across all models. Topographic attributes, soil properties, and climatic variables influenced phosphorus dynamics. Alluvial had the highest PUE due to favorable soil texture, while karst had lower efficiency due to phosphorus immobilization in carbonate-rich soils. Volcanic exhibited variable phosphorus availability. Despite weak correlations between environmental variables and phosphorus parameters, soil texture and slope were key determinants. Integration remote sensing and Random Forest model demonstrated high predictive accuracy and proving its effectiveness in estimating SAP, TPB and PUE.