Geospatial and Machine Learning Series

A five-phase, hands-on progression through a modern geospatial/ML tool stack, coordinate systems and spatial indexing, geospatial deep learning, Earth observation foundation models, spatial indexing/visualization, and natural-language search over imagery.

Practice project

Colorado Front Range Fire & Water Explorer public, non-Tribal data throughout, so this is a clean sandbox to build skill before deciding whether or how any of it ports back to CARE-governed work later. Suggested data:

• Marshall Fire (Dec 2021, Boulder County) burn scar and recovery
• Front Range snowpack/drought (SNOTEL stations, Colorado River headwaters)
• USGS stream gauges for a Front Range watershed of your choice

Structure

Phase	Folder	Tools	What you build
1	01-foundations-pyproj-rtree/	pyproj, rtree	Reproject two mismatched CO datasets by hand; build a spatial index over watershed boundaries
2	02-pytorch-torchgeo/	PyTorch, TorchGeo	Load a CRS-aware raster dataset and run a pretrained model on it
3	03-earth-embeddings/	AlphaEarth, GeoTessera	Train a small classifier on embeddings with a handful of hand-labeled points; compare open vs. proprietary embeddings on the same task
4	04-indexing-visualization/	H3, pydeck, Datashader	Bin embedding output into hex cells, render it interactively, then render full-resolution data at scale
5	05-semantic-search/	OpenCLIP / RemoteCLIP	Query your imagery in plain language

Each notebook has a Goal, an Exercise/Milestone, and a Check yourself prompt at the end. The code cells are stubs (# TODO) for you to fill in, not solutions.

Setup

python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt

Some phases need extra one-time setup:

• Phase 3 (AlphaEarth): requires a Google Earth Engine account ee.Authenticate() before first use.
• Phase 3 (GeoTessera): check coverage for your chosen area/year before pulling embeddings as coverage isn't automatically global.
• Phase 5 (RemoteCLIP): download the pretrained checkpoint from the RemoteCLIP repo, it loads directly into an OpenCLIP model.

Suggested sequencing

1. Foundations (pyproj, rtree)
2. PyTorch and TorchGeo
3. Earth embeddings (AlphaEarth, then GeoTessera for comparison)
4. Indexing and visualization (H3 to pydeck, then Datashader once data is large)
5. Semantic search (OpenCLIP/RemoteCLIP) capstone