AI.Geolabs · Premium GeoAI course
Deep Learning for Building Footprint Mapping
A convincing building mask is not enough. Learn to judge whether your footprint predictions are useful and what needs fixing before you use them.
Move beyond running a segmentation notebook. Learn how training labels, spatial evaluation, model choice, and prediction review affect the building footprints you deliver.
Hands-on exercises using 20 cm aerial imagery, QGIS, Python, and TorchGeo.
The problem behind the prediction
The notebook runs. The map still needs your judgment.
Very high resolution imagery reveals individual buildings in detail. It also reveals shadows, vegetation, complex roofs, and surrounding objects that can make extraction difficult.
You can spend hours digitizing buildings by hand or train a model and obtain a plausible-looking mask. But missing buildings, false detections, and irregular boundaries can remain. Once the mask becomes polygons, those mistakes become features in your GIS.
“It looks right. Is it?”
A clean-looking map can hide missed buildings and false detections. You need reference data and error inspection to assess its quality.
“Why did performance drop?”
Results on nearby test samples can give an optimistic picture. Applying the model elsewhere raises questions about imagery compatibility and transfer.
“Can I use these polygons?”
A polygon export is a useful step. It still needs review before supporting planning, infrastructure, or disaster assessment work.
What you will gain
Make better decisions about the footprints you produce.
The goal is to help you understand your results, assess their limitations, and decide what to review next.
Trust your evaluation
Know what your accuracy score actually supports.
Build spatial training, validation, and test splits. Select a classification threshold using validation data, and keep model confidence separate from measured accuracy.
Why it matters: You can make a more informed assessment of performance instead of relying on an impressive number alone.
Choose with evidence
Compare models without changing the rules.
Train U-Net and DeepLabV3+ using the same data and splits. Examine their results under consistent evaluation conditions.
Why it matters: You can assess the practical differences between the models for the course imagery.
Inspect the mistakes
Look more closely at missed and false detections.
Use region-targeted Grad-CAM to inspect model responses around detected, missed, and falsely detected buildings.
Why it matters: You gain another way to investigate predictions. Grad-CAM supports inspection; it does not prove a prediction is correct.
Apply and review
Take a trained model beyond its training tile.
Check new imagery for compatibility, predict with overlapping windows, and export georeferenced products plus low-confidence polygons for review.
Why it matters: You practice applying the model while retaining a clear role for quality checks and human judgment.
Beyond generated code
You can get segmentation code for free. What will you do with its results?
Free tutorials and AI assistants can help you get started. The harder questions arrive when the output looks plausible, the evaluation seems strong, or a new image gives unexpected results.
This course makes those decisions part of the learning: preparing consistent labels, comparing models fairly, choosing a threshold, inspecting errors, and reviewing exported footprints.
From “the code produced a map” to “I understand how this map was produced, how it was evaluated, and what still needs review.”
That is the course’s practical value: a worked mapping process that connects technical choices to the outputs you assess.
What the hands-on work gives you
Mapping products and the context to assess them.
Work with outputs you can inspect in a GIS, alongside the evaluation and interpretation needed to understand their limitations.
Products you will produce
- Reference building polygons and training masks.
- Building probability rasters and binary masks.
- Georeferenced candidate footprint polygons.
- Predictions on a compatible new tile.
- A list of low-confidence polygons for review.
Evidence you will examine
- U-Net and DeepLabV3+ results using the same spatial splits.
- A classification threshold chosen using validation data.
- Detected, missed, and falsely detected buildings.
- Region-targeted Grad-CAM explanations.
- Compatibility checks before applying the trained model.
High confidence does not guarantee correctness. Low-confidence flags help direct attention, but they do not identify every error. Predictions on unlabelled imagery remain candidate footprints until they are assessed and reviewed.
Is this course for you?
For analysts who need to understand the map they deliver.
This course is for GIS and remote sensing analysts, urban planners, and researchers who want practical building extraction skills with attention to evaluation and review.
It is particularly relevant if you have tried a segmentation notebook and now need to compare models, inspect mistakes, or apply a trained model to another image.
Recommended preparation
Basic familiarity with Python, QGIS, and raster and vector data will help you follow the exercises.
Plan for a GPU-enabled training environment, such as Google Colab. Training time depends on the resources available.
Imagery scope: The exercises use 20 cm aerial imagery. Applying the approach to satellite imagery or imagery with different characteristics requires compatibility checks and fresh evaluation.
Course coverage
The tools behind those decisions.
See the technical coverage
Prepare consistent building polygons in QGIS. Train U-Net with an ImageNet-pretrained ResNet34 encoder using TorchGeo, create reference masks and spatial splits, and choose a threshold using validation data. Train DeepLabV3+ with a ResNet50 encoder on the same data and splits. Inspect selected predictions with region-targeted Grad-CAM, then apply a trained model to a compatible new tile and export products for review. Each lab saves outputs needed by subsequent labs.
AI.Geolabs premium membership
Learn to assess the footprints—not just generate them.
Explore premium membership options for Deep Learning for Building Footprint Mapping. Check current pricing, course availability, and access terms on the membership plans page.
Already a premium member? Sign in to your AI.Geolabs account to check your course access.Before you join
A few practical questions
I can already generate code with AI. Why take this course?
The course provides a connected example of the decisions around the code: label preparation, spatial evaluation, model comparison, threshold selection, prediction inspection, and review of GIS outputs. Its value is in understanding and practicing that process.
Does this automate all building digitizing and review?
No. Reference digitizing is part of the training process, and model predictions still require assessment. The course teaches automated extraction alongside evaluation and review.
Will the model work on my own aerial or satellite imagery?
The final lab applies a trained model to a new tile without retraining, after compatibility checks. Different resolution, image characteristics, or building types can affect performance. Results on your own imagery need fresh assessment and may require additional training.
Will I produce GIS polygons?
Yes. The workflow exports georeferenced building polygons as well as probability rasters and binary masks. Polygons predicted on unlabelled imagery are candidates for review.
Where can I check membership pricing and access?
Visit the AI.Geolabs membership plans page for current pricing, course availability, and access terms.