International Journal of Science, Engineering and TechnologyAuthors: Manjusha Ipar, Suhani Jadhav, Samruddhi Kunde, Janhavi Landge, Mr. Sandeep Girijashankar Shukla
Abstract: River pollution is monitored almost exclusively through sparse in-situ sampling stations, leaving the large majority of a river network unobserved at any instant. This paper presents RiverWatch AI, a Windows desktop decision-support system that converts open multispectral satellite imagery into georeferenced, per-pixel pollution-risk assessments and an inventory of discrete hotspots for a user-selected river reach. The system implements a seven-stage geospatial pipeline — cloud screening, median compositing, water-index computation (NDWI, MNDWI, NDTI, NDVI), morphological water and channel extraction, a deterministic turbidity–sediment risk blend, connected-component hotspot detection, and polygon/CSV/PDF export — hosted in a responsive PySide6 interface whose heavy processing is dispatched to background threads. Rather than asserting an untrained neural network, the deployed estimator is a transparent, parameter-free spectral rule whose parameters, weaknesses and calibration targets are quantified. On a 900-pixel Sentinel-2 test scene the pipeline delineates 20,867 channel pixels (23.16 km²), flags 23.03 % of the channel as affected, resolves five hotspots with mean indicator confidence 0.912, and completes in 0.43 s. On a 900-sample labelled benchmark with 5-fold cross-validation, a Random Forest over the 13-D band-plus-index feature stack reaches 0.940 accuracy and 0.866 macro IoU, against 0.547 / 0.246 for the tuned index rule and 0.844 / 0.693 for an index-only Random Forest. The reported accuracy, precision, recall, F1, IoU, confusion matrices, affected-area statistics, hotspot tables and export artefacts are fully specified so that the system can be reproduced and independently verified.
DOI:
