Enhancing aridity index assessment in Pakistan's dryland ecosystems: A machine learning approach integrating remote sensing and seasonal lag effects

Abstract

Dryland ecosystems are highly vulnerable to increased aridity, thus exacerbating the drought stress. From this perspective, our study aimed to evaluate the aridity index (AI) and Standardized Precipitation Index at a three-month scale (SPI-3) across three arid stations of Pakistan from 1990 to 2023. Seven remote sensing indices were used as covariates with SPI-3 and mean temperature for enhanced prediction. Four well-optimized machine learning models were employed on seasonally decomposed time series. Mann-Kendall and Sen's slope analysis revealed a significant (p < 0.001) increasing trend of AI and SPI-3 values, indicating a comparatively lower aridity in recent years. It was consistent with the increasing trend of NDVI with Sen's slope range from 0.0002 to 0.003. Cross correlation showed a seasonal effect of biophysical indicators on AI with a positive correlation of r = 0.4 with NDVI and r = 0.6 with NDWI at lag 0, indicating a late lag effect. Furthermore, machine learning prediction of AI with a three-month lag size revealed an outperformance of Gradient Boosting Regression with R² = 0.806 and RMSE = 0.076, followed by Random Forest with R² = 0.732 and RMSE = 0.089. The Dry Barren Soil Index (DBSI), NDWI, and SPI-3 gained high feature importance in the highly performed model. Our study highlights the significance of temporal and seasonal relationships of aridity and biophysical indicators in dryland ecosystems, informing region-specific land and water resource management policies to mitigate hydroclimatic extremes.