Status: published
This paper presents a custom deep learning based land-cover classification system developed using Sentinel-2 satellite imagery and a trained U-Net style semantic segmentation model. In this project, satellite image patches and corresponding land-cover masks were prepared for the selected study area. The model was trained to generate pixel-wise land-cover predictions instead of only assigning one label to an entire image. The implemented system segments land-cover regions into water bodies, forest, agriculture, urban area, barren land, and grassland. Unlike approaches that simply display a pre-classified global product, the proposed work follows an end-to-end custom segmentation methodology: data collection, mask augmentation, preparation, encoder-decoder preprocessing, feature learning, validation, prediction-mask generation, database storage, and dashboard-based presentation. The frontend provides an interactive land-cover interface with maps, class visualization, statistics, screenshots, and user interaction, while the backend manages authentication, API services, model integration, and persistent records. The U-Net model learns spatial features such as texture, shape, boundaries, vegetation patterns, water smoothness, and built-up structures directly from the prepared Sentinel-2 samples. Experimental evaluation is reported using segmentation metrics such as pixel accuracy, Intersection over Union, Dice coefficient, precision, recall, and F1score. The project demonstrates that a student-developed deep learning and web GIS pipeline can transform raw satellite imagery into an accessible land-cover analysis tool suitable for environmental monitoring, urban planning, smart agriculture, and future GIS integration.
Keywords: Land cover classification, Sentinel-2, U-Net, semantic segmentation, deep learning, manual mask labeling, satellite imagery, Leaflet, MongoDB, Express.js, remote sensing.