GeoVision
An end-to-end geospatial AI pipeline detecting glacial lakes, roads, and drainage from satellite imagery.
- Status
- Completed
- Date
- 2025
- Category
- Intelligence
Summary
An end-to-end geospatial AI pipeline that detects glacial lakes, road centerlines, and drainage networks from satellite imagery and exports GIS-ready outputs.
Description
GeoVision is a Python-based computer vision and geospatial processing system for automated feature extraction from multispectral raster data. It replaces manual mapping workflows by running preprocessing, model training, inference, and geospatial post-processing in one pipeline.
Architecture
- Preprocessing
- Model training
- Large-raster inference
- Geospatial vectorization
- Temporal change analysis
Technology
- Python
- PyTorch
- segmentation-models-pytorch
- Albumentations
- Rasterio
- GeoPandas
- Streamlit
Features
- Automated geospatial feature detection pipeline for satellite imagery, covering glacial lakes, road centerlines, and drainage networks in a single workflow
- Multi-stage CLI execution for preprocessing, model training, inference, and end-to-end runs with configurable YAML-driven settings
- Deep-learning segmentation stack using U-Net/DeepLab-style models with augmentation, validation, early stopping, and checkpointed model saving
- Large-raster inference engine with tiling and overlap blending, plus feature-specific post-processing for cleaner masks and network continuity
- Geospatial vectorization workflow that converts predictions into polygons/lines/points, exports GeoJSON outputs, and computes area/length statistics
- Temporal change analysis utilities for multi-date imagery to track feature growth, shrinkage, emergence, and overall trend summaries
Links
Images
Machine-readable: /projects.json · /llms.txt · /llms-full.txt