urban_segmentation

Urban segmentation on multi channel satellite images to identify Residential and Industrial built up areas

MLDS MSc Final Project - Reichman University 2022

Intro

The repo contains the code implementation and trained models for the pipeline described above. The model was trained on 700 Sentinel2 GeoTiff images and 700 matching ESM masks A sample input image (before and after preprocessing) and matching ESM mask:

Our model scored a Dice score of 0.83 on test images and was successfully used to predict images on another part of the world:

Code

See the notebooks folder for all relevant notebooks:

• NB0 - Download new images from Google Earth Engine
• NB1 - Preprocessing and inference only
• NB2 - EDA and Preprocessing for Training
• NB3 - Minimal Preprocessing for Training or Inference
• NB4 - Training a new UNET segmentation model
• NB5 - Training a new DeepLabv3 segmentation model
• NB6 - Inference only
• NB7 - Training results analysis

Note: here is a link to a Youtube playlist with some "walkthrough" videos for the download and inference parts click here

How to use this repo

Downloading new Satellite images for prediction

• Use NB0 to download new images, you will need a GEE account and a Google Drive with free space for storing the images.
• The output will be a directory in google drive with your downloaded images.

Inference

• Use NB1 to preprocess and predict new masks for your downloaded images
• Or if you prefer you can preprocess using NB3 and then do Inference separately using NB6
• The output will be a directory with the predicted masks

Training a new model for your data

• Use NB2 (or NB3) for Preprocessing and NB4 (or NB5) for Training
• Use NB7 for analyzing
• The output will be a PyTorch model, and a directory with the predicted masks used for testing