Comparative evaluation of random forest, GRU, and transformer models for soil moisture prediction using reanalysis and meteorological time series data

Abstract

Soil moisture is an important indicator influenced by climate change and water resource availability, with significant impacts across various sectors, particularly agriculture. Limited continuous observational data necessitate historical data-driven approaches for accurate soil moisture prediction. This study aims to comparatively evaluate the performance of Random Forest, Gated Recurrent Unit (GRU), and Transformer models in soil moisture prediction using time series data based on a combination of NASA POWER reanalysis and BMKG meteorological data for the period 2020–2025. The methodology involves data acquisition, preparation, preprocessing (train–test split, min-max normalization, and windowing), model training, and performance evaluation based on MAE, RMSE, and R². The results show that the Random Forest model with a window size of 21 days achieves the best performance, yielding an MAE of 0.0396, an RMSE of 0.0534, and an R² of 0.8585. The Random Forest model produces predictions closest to the actual values and demonstrates better stability in capturing time series patterns, outperforming the GRU and Transformer models in soil moisture prediction using integrated global and local time series data