Predictive Quality Management in Nepalese Integrated Machine Learning and Geospatial Modelling for Urban Flood-Risk Prediction and LULC Forecasting In Nashik City, India

Abstract

Urban flood risk in fast-growing river cities is controlled jointly by land-cover transition, terrain, rainfall, drainage persistence, and hydrological data quality. This paper restructures the Nashik City thesis evidence into an integrated machine-learning and geospatial framework. Multi-epoch land-use/land-cover (LULC) maps for 1980-2024, a 2030 scenario, terrain derivatives, precipitation, land-surface temperature, runoff indicators, and 1980-2023 discharge records were analysed. K-nearest neighbours (KNN) provide a transparent classification benchmark, a shallow artificial neural network (ANN) supports nonlinear LULC projection, and a time-aware regression workflow produces discharge scenarios. Built-up cover rises from 3.80% in 1980 to 26.03% in 2024 and 29.12% in 2030, while trees decline from 9.53% to 3.01%. KNN reports 88.435% ten-fold mean accuracy at K=5 and 89.376% at K=31. A separate five-class validation reaches 77.03%, with water and trees recognised well but built-up land poorly detected. ANN validation accuracy increases from about 70.3% to 77.3%. Forecasts reproduce monsoon seasonality and report a 6,820 m3/s peak on 19 July 2025. Negative gauge values, gaps, class imbalance, absent uncertainty intervals, and unreported hydraulic calibration restrict operational use. The resulting 4D architecture links geospatial products to explicit historical and forecast dates for auditable planning support.