Accurate land use and land cover (LULC) classification combined with real-time change monitoring is fundamental to environmental stewardship, agricultural planning, and sustainable urban development. This paper presents GeoVision, a deep learning- enabled geospatial analysis framework that integrates EfficientNetB3 transfer learning for LULC classification, multi-spectral Change Vector Analysis (CVA) for temporal change detection, and a composite weather-aware risk assessment engine. The classification model was trained and evaluated on the EuroSAT benchmark dataset comprising 27,000 Sentinel-2 tile images across ten land cover categories, achieving an overall accuracy of 94.2%, a top-3 accuracy of 98.7%, and a macro-averaged F1-score of 93.5%. Spectral indices—NDVI, NDMI, and NDBI—were computed over Landsat 8/9 and Sentinel-2 imagery retrieved dynamically through Google Earth Engine (GEE) to quantify vegetation dynamics, moisture change, and urban expansion. CVA aggregates per-index deltas into a unified change magnitude to classify regions as stable, moderately altered, or significantly changed. The system further incorporates real-time meteorological data from the OpenWeather API, enabling a multi-factor risk score computed over temperature extremes, humidity thresholds, wind velocity, and precipitation intensity. GeoVision is deployed as a fully interactive Streamlit web application with Folium-based geospatial visualization. Experimental results demonstrate a mean NDVI change detection error of ±4.8%, urban expansion estimation error of ±7.6%, and consistent risk score calibration against known environmental events. GeoVision addresses the critical research gap in which existing LULC systems offer either classification alone or limited change detection, without the unified integration of real-time satellite retrieval, weather coupling, and risk quantification that operational environmental monitoring demands.
Babu et al. (Sun,) studied this question.