Awesome
Infrastructure Mapping and Monitoring in Deserts with Sentinel-1
This project implements a methodology for infrastructure mapping and monitoring in desert regions using Sentinel-1 SAR data. The methodology consists of the following steps:
- Extract a time series of 6 consecutively acquired Sentinel-1 GRD images over an area of interest. Here the data is obtained through Creodias. The areas of interest can be found in AOI folder
(see AOI folder)
- Apply the following preprocessing steps automatically through shell scripts and GPT (command line version of SNAP software):
- Calibration to Sigma0 (VV and VH)
- Stacking and multitemporal speckle filtering
- Terrain Correction (using SRTM DEM) and conversion to decibel
- Coherence generation for consecutively acquired pairs, then averaging all
(see SentProc folder)
- Apply the following Deep Learning workflow:
- Create mask of known infrastructure from Open Street Map (rasterise OSM vectors)
- Extract patches from areas of Sentinel-1 data covered by mask
- Divide patches between train, validation and test data
- Augment training data
- Train U-Net model for image segmentation
- Extract patches over entire Sentinel-1 data, apply the model to these patches
- Reconstruct image from model output, and convert raster to vector
(see DeepLearning folder)