Geoai With Python

GeoAI with Python

Source: Qiusheng Wu, GeoAI with Python: A Practical Guide to Open-Source Geospatial AI
Edition context: official web book, 430-page print edition, 23 chapters
Generated: 2026-08-01

How to use this skill

• Ask for a workflow: “map water bodies from Sentinel-2”, “train a building detector”, or “compare two dates”. I will choose a task, data contract, baseline, and validation plan.
• Ask for a chapter or topic: ch09, STAC, SAM, QGIS, tile overlap, or embeddings. I will read the linked chapter before answering.
• Ask to diagnose an error. I will inspect the first failed boundary—environment, data shape/CRS, tiling, model configuration, inference, or export—before suggesting a rerun.
• If the task depends on current package APIs, data endpoints, model access, or GPU/CUDA compatibility, verify the live official documentation before executing.

Operating loop: goal → evidence → adjustment

1. Define the target and success evidence. State the geographic area, dates, input sensors/bands, target object/value, output format, acceptable quality, and compute limit.
2. Inspect before modeling. Confirm raster shape, band count/order, dtype, nodata, CRS, resolution, bounds, temporal coverage, and vector label CRS/geometry. Never assume a downloaded file is model-ready.
3. Choose the smallest useful method. Use pre-trained/foundation models for rapid extraction; use supervised training when labels, repeatability, or domain adaptation justify it; use classical differencing as a change-detection baseline.
4. Prepare spatially. Reproject labels to the raster CRS, align grids, rasterize or vectorize deliberately, tile large rasters, and split by geographic scene/region where possible to avoid spatial leakage.
5. Run a small baseline first. Use one tile or a small sample, low epochs, and saved outputs. Check that the model learns and that predictions overlap the source imagery before scaling up.
6. Validate the map, not only the metric. Inspect overlays and failure cases; report task metrics plus spatial artifacts, edge effects, small-object loss, class imbalance, and transfer limitations.
7. Export reproducibly. Keep the model/config/data provenance, save georeferenced rasters or vectors, add geometric properties, filter artifacts with explicit thresholds, and preserve intermediate masks/probabilities when they explain decisions.

Task selection

Non-negotiable data contract

• Spatial: inputs and labels cover the same area; CRS is explicit; vector operations use a projected CRS when measuring area/length.
• Raster: band order and sensor-specific scaling are recorded; nodata is handled; multispectral channel count matches the model.
• Tiles: choose tile_size from object scale and GPU memory; use stride < tile_size for overlap; inspect edge artifacts and duplicate detections.
• Splits: make train/validation/test spatially independent where the scene permits; do not split adjacent chips from one image randomly and call that geographic generalization.
• Outputs: retain probability/score rasters for threshold analysis; vectorize only after mask cleanup; filter by area/shape only with a documented reason.

Core frameworks and reusable methods

The GeoAI pipeline

Treat every project as acquire → inspect → prepare → learn/infer → evaluate → spatialize → validate → publish. A failure in a later stage should send you back to the earliest violated contract, not trigger blind retries.

Raster–label alignment

Use the raster as the reference grid. Reproject vector labels, rasterize with the reference transform, and verify an overlay before tiling. For masks, distinguish background, nodata, and “unknown”; they are not interchangeable.

Patch-based scale-out

For large imagery, tile with overlap, infer batches, and stitch in the source CRS. Increase overlap when objects are cut at edges; reduce it only after checking seam quality and memory. Use a smaller window or batch size before changing the model when CUDA memory fails.

Model choice by evidence

Prefer a pre-trained model when the target is visually recognizable and the goal is a fast baseline. Prefer supervised training when the target definition, sensor, geography, or output precision differs from the pre-training domain. Compare architectures only after the data pipeline and validation split are stable.

Prompt → mask → geometry

For SAM, choose the weakest prompt that identifies the intended object, inspect the mask, save scores/IDs, then convert and regularize polygons. A prompt is not ground truth; measure precision/recall or sample-based accuracy before operational use.

Continuous prediction discipline

For pixel regression, constrain valid ranges, mask invalid target pixels, evaluate with MAE/RMSE/R² plus residual maps, and compare predictions against a held-out date or region. A visually smooth raster can still be systematically biased.

Chapter index

Topic index

• COG / Cloud Optimized GeoTIFF → ch03, ch05
• CRS / reprojection / alignment → ch03, ch06
• CUDA / GPU / memory → ch02, ch09, ch14, ch17
• embeddings / TESSERA / AlphaEarth → ch16
• image recognition → ch01, ch07
• image translation / super-resolution → ch11
• instance segmentation / Mask R-CNN → ch10, ch23
• object detection / COCO / mAP → ch08, ch18, ch20
• QGIS → ch17–ch23
• SAM / prompts / regularization → ch14, ch21
• semantic segmentation / U-Net → ch09, ch19, ch22
• STAC / Planetary Computer → ch04, ch05
• training chips / tiling / label quality → ch06
• VLM / Moondream / CLIP → ch15, ch20
• water / cloud / land cover → ch09, ch19, ch22

Supporting files

• glossary.md — concise definitions and chapter pointers.
• patterns.md — executable workflow patterns and trade-offs.
• cheatsheet.md — one-page decisions, parameters, and checks.
• references/source-map.md — official sources, code/data links, and license boundaries.
• references/qgis-training-inference.md — supplementary QGIS page present in the repository but not in the 23-chapter table of contents.

Scope and limits

This skill synthesizes the official book and repository into an execution guide; it is not a substitute for current package documentation. APIs, model access, datasets, and QGIS plugin behavior can change. Verify live official sources before production use, and do not treat model predictions as authoritative measurements without domain validation.