Enhancing assimilated soil moisture prediction from environmental data using advanced machine learning

Soil moisture is a vital component of terrestrial water cycles, shaping land atmosphere interactions and its accurate predictions in data scarce regions is a major challenge. Data driven machine learning models enhance the prediction accuracy of assimilated soil moisture in arid regions. From this perspective, our study aimed to enhance the prediction accuracy of top layer soil moisture (0–10 cm) in arid irrigated and rainfed regions of Pakistan using multisource environmental data, using advanced machine learning models. The t-Distributed Stochastic Neighbor Embedding (t-SNE) dimension reduction approach was integrated into three machine learning (ML) models including Random Forest (RF), Gradient Boosting Regression (GBR), and Artificial Neural Network (ANN). Exploratory analysis revealed significantly higher soil moisture in rainfed (0.245 m³/m³) than irrigated (0.175 m³/m³), with strong correlations to net longwave radiation, humidity, and latent heat flux. Principal component analysis showed high dimensionality (PC1: 55.9%, PC2: 27.3%). ML output revealed that 2D-t-SNE + GBR achieved high accuracy (R2 = 0.889) Vs GBR without t-SNE (R2 = 0.845) in irrigated region. In the rainfed region, 3D-t-SNE + GBR achieved highest R2 = 0.754. The study provides more reliable soil moisture estimation supporting drought early warning system and climate resilient agriculture.