Ultimate Goal

Leverage deep learning techniques for satellite imagery analysis to create meaningful information for advocacy groups about the progression of systematic land dispossession in the West Bank.

Intermediate Goals

1. Learn what Python libraries are used to process, shape, and visualize satellite images for deep learning models.
2. General acquaintance of different pre-trained models for “large vision” tasks and the datasets they were trained on.
3. Practice different deep learning tasks on satellite images e.g classification, object detection, segmentation, anomaly detection w/ auto encoders.

Data

While typical images hold 3-4 channels (RGBA) satellite images can hold dozens of channels that hold different wavelengths not visible to humans. For practice I’ll concentrate on sentinel-2 satellite images which are commonly used for agriculture analysis tasks and resemble regular aerial images.

Practice datasets:

• Github Repo: awesome satellite imagery
• Github Repo: awesome remote sensing change detection

Actual dataset:

• TBD after discussion with SITU and visit to Israel. There is definitely at least a limited number of annotated images. Preparing for less data and willing to be pleasantly surprised.

Models

Meta’s Segment Anything Model has a geospatial version trained on RGB channels of satellite images. This is GREAT for object detection (like counting buildings), but might not be enough for land cover. That’s because models usually determine the type of crop according to invisible light that bounces off the fields, which is measured in the other channels of the image.

There is also a list of multi-spectral models recently trained and tested by the University of Berlin:

https://huggingface.co/BIFOLD-BigEarthNetv2-0

Stages / Exercises

Research and mapping of tutorials, repos, papers, libraries, and models.

Preprocessing satellite images tutorial & segmentation tutorial