Land Cover Mapping Using Deep Learning: A Remote Sensing Framework for Carbon Stock Assessment
Contributors
Vijayalakshmi V
Raja Sarath Kumar boddu
Keywords
Proceeding
Track
General Track
License
Copyright (c) 2026 Sustainable Global Societies Initiative

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Abstract
Accurate spatial estimation of terrestrial carbon stocks is essential for understanding ecosystem carbon dynamics, supporting climate-change mitigation, and developing sustainable land-management strategies. Conventional field-based biomass inventories provide reliable measurements but are spatially limited, labour-intensive, and difficult to scale over large geographic regions. Satellite remote sensing provides an effective means of obtaining spatially continuous observations, while recent advances in deep learning enable the extraction of complex spectral and spatial relationships from remotely sensed imagery. This study proposes an integrated deep-learning framework for land-cover mapping and spatial carbon-stock estimation using multispectral remote sensing data. Landsat 8/9 spectral bands are combined with vegetation and land-surface indices, including the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-up Index (NDBI), and Soil-Adjusted Vegetation Index (SAVI). The resulting feature space is used for pixel-level land-cover classification and continuous above-ground biomass (AGB) estimation. Land-cover information is subsequently integrated with spatial biomass estimates to derive carbon density and total carbon stock. Unlike conventional LULC-based approaches that assign fixed carbon coefficients to land-cover classes, the proposed framework treats biomass as a continuous spatial variable, thereby accounting for within-class variability in vegetation structure. The framework can additionally incorporate Sentinel-1 synthetic aperture radar (SAR), Sentinel-2 optical observations, digital elevation models, and texture features to improve biomass prediction. Land-cover classification is evaluated using overall accuracy, precision, recall, F1-score, and Intersection over Union (IoU), while biomass estimation is evaluated using the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and relative error. The resulting carbon-stock maps provide spatially explicit information on carbon storage and enable the identification of carbon-rich regions and potential carbon-loss hotspots associated with land-cover change. The proposed framework provides a scalable approach for integrating deep learning and multi-source satellite observations for ecosystem carbon assessment.