Scholarlink

<p>Accurate high-resolution rainfall information is essential for hydrological modeling land use planning and climate impact assessment, particularly in data-scarce regions. However, widely used satellite-based precipitation products, such as the Integrated Multi-Satellite Retrievals for GPM IMERG, are limited by their coarse spatial resolution, which restricts their direct application at local scales. This study presents a machine learning based framework to downscale monthly IMERG rainfall data from 10&#xa0;km to 500&#xa0;m spatial resolution by integrating the OPTRAM optical soil moisture index with key environmental variables derived from freely available remote sensing data. The framework was implemented for the agriculturally important Sahiwal district of Pakistan for the years 2019 and 2020. Two machine learning models, Random Forest and Epsilon Support Vector Regression, were evaluated using ground-based rain gauge observations. Environmental predictors including Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), and Normalized Difference Water Index (NDWI) derived from Sentinel 2 imagery at 10&#xa0;m spatial resolution, together with Enhanced Vegetation Index (EVI) and Land Surface Temperature (LST) obtained from MODIS products at 1&#xa0;km resolution, were integrated with OPTRAM based soil moisture estimates to guide the downscaling process. The Random Forest model demonstrated superior performance, achieving coefficients of determination R² of 0.98 for 2019 and 0.89 for 2020 with corresponding RMSE values of 7.09&#xa0;mm per month and 12.25&#xa0;mm per month, respectively and lower mean absolute errors compared to the Epsilon SVR model. Residual correction further improved the agreement between downscaled and observed rainfall. Variable importance analysis identified OPTRAM-derived soil moisture and land surface temperature as the most influential predictors in the downscaling process. The results demonstrate that integrating optically derived soil moisture with environmental covariates substantially enhances the spatial detail and accuracy of satellite rainfall estimates. The proposed scalable framework, implemented on the Google Earth Engine platform, offers a practical solution for high-resolution rainfall estimation, supporting hydrological modelling, climate impact studies, and precision agriculture in regions with limited in situ observations.</p> Graphical Abstract <p></p> <p>The graphical abstract presents a visual summary that intends to gives a concise overview of the study’s core findings and methodologies. This abstract provides a robust, scalable framework developed in this study to enhance applicability of satellite precipitation products for fine-scale climate impact assessments, hydrological modeling, and precision agriculture, particularly in data-scarce environments. The study emphasized the need for accurate fine-scale rainfall information and its importance in hydro-meteorological applications under changing climate. It showed enhanced utility of coarse-resolution global satellite precipitation products in regions where localized climate information is critical for water-resource management and land-use planning. The graphics summarizes the workflow of input datasets employed in the machine learning–based downscaling approach. Coarse-resolution IMERG rainfall (10&#xa0;km) is combined with key environmental covariates, including OPTRAM-derived soil moisture, NDVI, land surface temperature (LST), and NDWI. These variables, representing optical, thermal, and hydrological surface conditions, are integrated within a machine learning model designed to capture the spatial and environmental drivers of rainfall variability. The central model block symbolizes this computational integration and the predictive relationships learned across multiple datasets. The study generated high-resolution (500&#xa0;m) rainfall fields and identified the most influential predictors to substantially improve spatial rainfall representation in the downscaling process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Soil-Moisture-Constrained ML Framework for High-Resolution Rainfall Estimation from Satellite Data

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Syed Muhammad Talha,
  • Mujtaba Hassan,
  • Ahmad Bilal,
  • Farrukh A. Chishtie,
  • Bashir Ahmad,
  • Asim Qadeer,
  • Muhammad Zia UR Rahman Hashmi,
  • Imran Shahid

摘要

Accurate high-resolution rainfall information is essential for hydrological modeling land use planning and climate impact assessment, particularly in data-scarce regions. However, widely used satellite-based precipitation products, such as the Integrated Multi-Satellite Retrievals for GPM IMERG, are limited by their coarse spatial resolution, which restricts their direct application at local scales. This study presents a machine learning based framework to downscale monthly IMERG rainfall data from 10 km to 500 m spatial resolution by integrating the OPTRAM optical soil moisture index with key environmental variables derived from freely available remote sensing data. The framework was implemented for the agriculturally important Sahiwal district of Pakistan for the years 2019 and 2020. Two machine learning models, Random Forest and Epsilon Support Vector Regression, were evaluated using ground-based rain gauge observations. Environmental predictors including Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), and Normalized Difference Water Index (NDWI) derived from Sentinel 2 imagery at 10 m spatial resolution, together with Enhanced Vegetation Index (EVI) and Land Surface Temperature (LST) obtained from MODIS products at 1 km resolution, were integrated with OPTRAM based soil moisture estimates to guide the downscaling process. The Random Forest model demonstrated superior performance, achieving coefficients of determination R² of 0.98 for 2019 and 0.89 for 2020 with corresponding RMSE values of 7.09 mm per month and 12.25 mm per month, respectively and lower mean absolute errors compared to the Epsilon SVR model. Residual correction further improved the agreement between downscaled and observed rainfall. Variable importance analysis identified OPTRAM-derived soil moisture and land surface temperature as the most influential predictors in the downscaling process. The results demonstrate that integrating optically derived soil moisture with environmental covariates substantially enhances the spatial detail and accuracy of satellite rainfall estimates. The proposed scalable framework, implemented on the Google Earth Engine platform, offers a practical solution for high-resolution rainfall estimation, supporting hydrological modelling, climate impact studies, and precision agriculture in regions with limited in situ observations.

Graphical Abstract

The graphical abstract presents a visual summary that intends to gives a concise overview of the study’s core findings and methodologies. This abstract provides a robust, scalable framework developed in this study to enhance applicability of satellite precipitation products for fine-scale climate impact assessments, hydrological modeling, and precision agriculture, particularly in data-scarce environments. The study emphasized the need for accurate fine-scale rainfall information and its importance in hydro-meteorological applications under changing climate. It showed enhanced utility of coarse-resolution global satellite precipitation products in regions where localized climate information is critical for water-resource management and land-use planning. The graphics summarizes the workflow of input datasets employed in the machine learning–based downscaling approach. Coarse-resolution IMERG rainfall (10 km) is combined with key environmental covariates, including OPTRAM-derived soil moisture, NDVI, land surface temperature (LST), and NDWI. These variables, representing optical, thermal, and hydrological surface conditions, are integrated within a machine learning model designed to capture the spatial and environmental drivers of rainfall variability. The central model block symbolizes this computational integration and the predictive relationships learned across multiple datasets. The study generated high-resolution (500 m) rainfall fields and identified the most influential predictors to substantially improve spatial rainfall representation in the downscaling process.