Who this is for

You work with rasters or remote sensing and want to cross into deep learning without hand-waving. Or you are a data scientist who knows neural networks and needs to understand why they behave differently on a grid of pixels with a coordinate system. This journey builds the models from first principles — a single convolution filter up to a U-Net — on real satellite, aerial, and LiDAR data.

Assumes: the Getting Started journey, or equivalent — comfortable Python, NumPy array thinking, and the geospatial stack at the level of the first Essentials book. No deep learning or PyTorch experience is assumed; that is what this journey is for.

By the end you can

• Search a public satellite archive, download exactly the bands you need, and convert them to surface reflectance correctly — no account, no API key.

• Build every component of a U-Net in PyTorch yourself, and say why a fully connected network is the wrong tool for a raster.

• Map cropland, detect change between two dates, segment buildings from aerial imagery, and separate buildings from trees with LiDAR — each evaluated on a geographic hold-out, not a shuffled split.

• Choose between architectures, and between training from scratch and a foundation model.

• Connect a language model to spatial tools and get an answer you can check.

How to follow this journey

• Go in order. Each step assumes the one before it.

• Start with step 1, then run Chapter 0 of the Satellite course — if your laptop reaches the public catalog, you have everything the journey needs.

• Learn more opens the course or book page; Start the course / Get the book takes you to Gumroad. Most items have a sample there, so you can check the fit first.

• Going deeper is optional and unordered — a shelf, not a route. Come back to it after the last step, or whenever a step raises a question.

The route

6 steps · roughly 6.5 hours of video plus one book

1. What Is GeoAI — and Why Does It Matter?

The bigger picture before the first tensor: what GeoAI is, where it already works, and where it still fails.

Read →   Free article

2. Satellite Data Science in Python

Sentinel-2 from scratch, in pure Python, on a normal laptop. A raster you build by hand, a real acquisition pipeline, three spectral indices proved as separators, a two-year data cube, and a land-cover map of Manhattan you trained — with every result interpreted and its limit stated out loud. This is the raster fluency every model below assumes.

Learn more →  ·  Start the course →   Course · ~2.5 h

3. Fundamentals of GeoAI: Deep Learning for Geospatial Analysis

A U-Net assembled from a single Conv2d filter, then put to work four times: crop mapping in Hungary, change detection with a Siamese network, building segmentation over Amsterdam, and buildings versus trees from Scottish LiDAR. You know what every block does before you train it.

Learn more →  ·  Start the course →   Course · ~3.5 h

4. Building Classification with GeoAI — Google AlphaEarth in Python

First contact with foundation-model embeddings: 64 bands per pixel, downloaded directly, joined to OpenStreetMap footprints. The step between the models you built and the ones you will borrow.

Watch →  ·  Read the write-up* →   Free video · 24 min

5. How I Built a GeoAI Agent That Answers Spatial Questions in Plain English

Five minutes on where the book ends: a language model connected to five spatial tools, asked about a city, and checked. Watch it before you commit to the 460 pages.

Watch →  ·  Read the write-up* →   Free video · 5 min

6. Geospatial Data Science Essentials: 101 Steps to GeoAI from Scratch

The course in print, and then the eight chapters no course covers: object detection, height regression, spatio-temporal forecasting, cloud gap-filling, foundation models compared, and the agent — built from scratch. Every model evaluated on regions it never saw.

Learn more →  ·  Get the book →   Book

Going deeper

Optional, any order.

Detecting Wildfire Damage with Python, Satellite Data & OSM

A complete satellite pipeline — Sentinel, NASA FIRMS, OSM, burn-ratio differencing — with no deep learning in it at all.

Watch →  ·  Write-up, part 1* →  ·  Part 2* →   Free video · 24 min

Google's New Rooftop Dataset + Python: 12 Cities Compared

Working with a satellite-derived machine-learning dataset, and why London's roofs are dark.

Watch →  ·  Read the write-up* →   Free video · 6 min

A Comprehensive List of Satellite Indices

The course teaches three indices properly; this is the reference for the other forty.

Read* →   Reference

The Public Geospatial Foundation Model Landscape in 2026

Seventeen models, who built them, and the numbers that matter.

Read* →   Reference

101 Essential GeoAI Concepts

The field's working vocabulary, one plain sentence each.

Read* →   Reference

30 Data Sources for GeoAI

Where to find the data for your own next project.

Read* →   Reference

10 Papers on Spatial Downscaling

Elective reading on sharpening coarse satellite products.

Read* →   Reading list

* Part of my premium Substack plan.