Drought index downscaling using AI-based ensemble technique and satellite data

Abstract

AbstractThis study introduces and validates an artificial intelligence (AI)–based downscaling method for Standardized Precipitation Indices (SPI) in the northwest of Iran, utilizing PERSSIAN-CDR data and MODIS-derived drought-dependent variables. The correlation between SPI and two drought-dependent variables at a spatial resolution of 0.25° from 2000 to 2015 served as the basis for predicting SPI values at a finer spatial resolution of 0.05° for the period spanning 2016 to 2021. Shallow AI models (Support Vector Regression, Adaptive Neural Fuzzy Inference System, Feedforward Neural Network) and the Long Short-Term Memory (LSTM) deep learning method are employed for downscaling, followed by an ensemble post-processing technique for shallow AI models. Validation against rain gauge data indicates that all methods improve SPI simulation compared to PERSIANN-CDR products. The ensemble technique excels by 20% and 25% in the training and test phases, respectively, achieving the mean Determination Coefficient (DC) score of 0.67 in the validation phase. Results suggest that the deep learning LSTM method is less suitable for limited observed data compared to ensemble techniques. Additionally, the proposed methodology successfully detects approximately 80% of drought conditions. Notably, SPI-6 outperforms other temporal scales. This study advances the understanding of AI-driven downscaling for SPI, emphasizing the efficacy of ensemble approaches and providing valuable insights for regions with limited observational data.