ANNOUNCEMENTS
Riverine sand, a critical component of fluvial ecosystems and a key resource for the construction industry, has increasingly come under environmental scrutiny due to rampant, unregulated mining. Traditional methods of monitoring sand deposits are constrained by scale, time, and resources, particularly across expansive and ecologically diverse regions such as India’s major river basins. This study develops a machine learning–based framework for the detection and mapping of riverine sand deposits using Sentinel-2 satellite imagery and hydrological modeling, applied to eight major Indian rivers: Ganga, Brahmaputra, Godavari, Krishna, Narmada, Tapi, Mahanadi, and Cauvery.
Using non-monsoon, cloud-free multispectral imagery and derived indices such as the Normalized Difference Sand Index (NDSI) and its enhanced variant (NDESI), the study trains and evaluates four classification models—Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and Convolutional Neural Networks (CNN). Extensive feature extraction, manual labeling, and validation were conducted across five geographically diverse test sites, augmented by cross-referencing with Dynamic World and SIS-DP datasets.
Among all models, Random Forest achieved the highest classification accuracy (92%), proving robust across varied terrain and hydrological conditions. Spatial outputs from the RF model reveal distinct patterns of sand distribution, with significant concentrations in the Ganga and Brahmaputra basins and more localized, regulated deposits in peninsular river systems. This scalable and repeatable methodology offers an efficient, accurate, and automated tool for environmental monitoring, aiding policymakers and local authorities in promoting sustainable sand resource management and curbing illegal extraction practices.
