GeoAI - From a raster you draw by hand to a model you trained

There is a version of GeoAI that starts by downloading a pretrained model and running it. It produces a map on the first afternoon, and it leaves you unable to say why the map is wrong when it is wrong. This GeoAI journey takes the other route: six steps, in which the first model you train arrives at step three; you learn how to build and troubleshoot models from scratch, and the models you borrow arrive only after the foundations are laid down properly. An extra afternoon, in return for being a conscious user rather than a button-presser.

The journey assumes the Getting Started journey or the equivalent — comfortable Python, NumPy array thinking, and the geospatial stack at the level of the first Essentials book. It does not assume any deep learning or PyTorch; that is what it is for.

Step 1–2: the framing, then the data

Step one is the article What Is GeoAI — and Why Does It Matter. The field's vocabulary borrows from remote sensing, deep learning, and classical GIS all at once, and a learner who cannot tell those three apart will not know which of their assumptions is breaking. The article's job is to say where AI has actually changed what is possible — floods, crops, cities, disasters — and where it still fails. From there we start where nearly every GeoAI pipeline starts: getting our hands on satellite data.

Satellite Data Science in Python has no deep learning in it, and that is its contribution. It builds the thing every model downstream assumes: clean, correct, well-understood rasters. You start with a 64×64 array that becomes a georeferenced GeoTIFF the moment you give it a transform and a CRS. Then a real pipeline — search a public catalogue by place, date, and cloud cover; pull only the bands you need; convert raw integers to surface reflectance correctly. Then, we move on to build three spectral indices from one arithmetic recipe and prove them as separators with distributions rather than pretty maps. A two-year data cube teaches the difference between seasonality and change. A land-cover classifier of Manhattan is trained from OpenStreetMap labels, and then a fourth class is added specifically to watch it fail at ten metres — because a sensor telling you its limit is a result, not a defeat.

Nothing in that course needs an account, an API key, or a GPU. Everything in the rest of the journey needs the habits it builds.

Step 3: building the network yourself

The course Fundamentals of GeoAI does not begin from a pretrained black box. It begins from a single convolution filter and assembles a U-Net from it — filters, an encoder block, a stacked encoder, a decoder with upsampling, skip connections — so that by the time you train it, you know what every block does and why it is there. The course is precise about why a fully connected network is the wrong tool for a 128×128×4 raster: 65,536 inputs, exploded parameters, and the spatial structure thrown away before learning starts.

Then the same architecture is put to work four times, on four real study areas: cropland in Hungary from Sentinel-2, mapped three ways so the methods can be compared honestly; change between April and July with a Siamese U-Net whose two encoders share weights; building segmentation over Amsterdam from high-resolution aerial imagery, with a spatial train/test split and IoU rather than a pretty overlay; and buildings versus trees from Scottish LiDAR, a distinction spectral data alone cannot make.

That phrase — spatial train/test split — is the one that separates GeoAI from generic machine learning, and the journey returns to it at every step. A random split flatters a spatial model, because nearby pixels leak. Every model here is evaluated on regions it never saw. When something goes wrong in your own work later — and it will — this is the first thing you will check, because you will have watched it go wrong on purpose.

Step 4–6: borrowing, then extending

Step four, Building Classification with Google AlphaEarth, is the first contact with a foundation model — 64 bands of learned embeddings per pixel, downloaded directly and joined to OpenStreetMap footprints. It sits after the from-scratch course on purpose: you can only judge what a foundation model gives you once you know what training one costs.

Step five is five minutes on where the book ends — a language model connected to five spatial tools, asked plain-English questions about a city, and checked. Step six is the book itself. 101 Steps to GeoAI from Scratch is the course in print plus the eight chapters no course covers: object detection with sliding-window inference, height regression that replaces LiDAR with spectral bands, spatio-temporal forecasting with a ConvLSTM, cloud gap-filling with an autoencoder, four foundation models compared on the same tasks, and the agent — all built from scratch in PyTorch, on real Sentinel-2, LiDAR, and aerial data, every model evaluated on geographically held-out regions.

What you will have

This journey teaches pixel-level segmentation and the full pipeline around it — data acquisition, labels, patching, spatial splits, training, evaluation, inference — and then shows you how to bring in models you did not train. What it hands you is the ability to read a segmentation map and say which of its pixels to trust, and to debug the pipeline when the answer is "not these." In a field where the model is no longer the bottleneck and the reference data is, that turns out to be the durable skill. No GPU, no cloud compute, no credit card: everything runs on a laptop.

Extra steps

After the core route, the shelf branches in three directions. Applied pipelines without deep learning: Detecting Wildfire Damage with Python, Satellite Data & OSM is a complete burn-scar analysis — Sentinel imagery, NASA FIRMS fire detections, OpenStreetMap buildings, burn-ratio differencing — and Google's Rooftop Dataset, 12 Cities Compared is a worked example of using a satellite-derived machine-learning product rather than building one. References for when the course's three indices are not enough: A Comprehensive List of Satellite Indices covers the other forty, 101 Essential GeoAI Concepts is the field's vocabulary in one plain sentence each, and 30 Data Sources for GeoAI is where to find data for your own next project. And two pieces for going further into models: The Public Geospatial Foundation Model Landscape 2026 surveys the seventeen public models behind the book's Chapter 13 decision framework, and 10 Papers on Spatial Downscaling is elective reading on sharpening coarse satellite products to higher resolution.