Land Cover Classification Using Multi-Source Satellite Data: Achieving 97% Accuracy via Sentinel-1 and Landsat 8 Fusion

This project fuses Sentinel-1 C-band SAR and Landsat 8 optical data, leveraging the Google Earth Engine platform and machine learning techniques to achieve 97% land cover classification accuracy in the Los Angeles area. The project is user-friendly; non-technical users can complete complex data analysis through an intuitive interface, supporting applications in urban planning, environmental monitoring, and other scenarios.

The project uses two complementary data sources:

• Sentinel-1 SAR: A radar satellite from the European Space Agency, with all-weather and day-night observation capabilities. Its C-band is suitable for monitoring vegetation and urban areas, with a 6-day revisit cycle;
• Landsat 8: A joint satellite by NASA/USGS, with multi-spectral (11 bands) and 30m resolution, 16-day revisit cycle, providing spectral features and surface temperature information. Fusing the two can leverage their respective advantages.

The technical architecture includes:

1. Data Fusion: Combining SAR's surface roughness information with optical spectral features, using GEE cloud computing;
2. ML Classification Workflow:
   • Preprocessing: SAR radiometric/geometric correction, optical atmospheric correction and cloud masking, spatiotemporal registration;
   • Feature Engineering: Extracting SAR backscatter coefficients, vegetation/water indices, texture features, and temporal features;
   • Model Training: Supporting Random Forest, SVM, neural networks, etc., with cross-validation to ensure generalization;
   • Result Output: Generating thematic maps, area statistics, and exporting formats like GeoTIFF/Shapefile.

Taking the Los Angeles area as the test site (with complex land cover types: built-up areas, vegetation, water bodies, bare land, etc.), a multi-level classification system is adopted. Accuracy verification results: overall accuracy exceeds 97%, Kappa coefficient >0.95, and user/mapping accuracy for all categories is higher than 90%, meeting application needs such as urban planning and disaster assessment.

The project lowers the threshold for remote sensing technology; non-professional users can complete analysis via a graphical interface. It uses the MIT open-source license, supporting algorithm improvements, multi-language localization, and GIS integration. It provides accurate land cover data for climate change research, aiding urban expansion monitoring, deforestation assessment, etc.

The team plans to launch enhanced features: 3D terrain overlay and temporal animation, automatic updates of real-time satellite data streams, integration of more deep learning models, cloud collaboration, mobile adaptation, etc., to continuously improve the tool's practicality and accessibility.