Fundamentals of GeoAI: Deep Learning for Geospatial Analysis

Whether you work in GIS, remote sensing, environmental science, or data science, deep learning is rapidly transforming how we analyze the world from above. This course gives you the practical foundation to work with GeoAI confidently — building real models on real data, from scratch.Across five hands-on modules, you will tackle the most important use cases in geospatial deep learning today: crop mapping from Sentinel-2 satellite imagery, temporal change detection using a Siamese U-Net, building segmentation from ultra-high-resolution aerial imagery, and multi-class urban segmentation from LiDAR elevation data.Every dataset in this course is real and freely available. Sentinel-2 imagery is downloaded directly from the AWS Earth Search STAC catalog. Aerial imagery comes from a Dutch government geoportal at 7.5cm resolution. LiDAR tiles are sourced from the Scottish Government open data portal. No synthetic data, no toy examples.Every model is built from scratch in PyTorch. You will implement single convolutional filters, build encoder and decoder blocks step by step, assemble complete U-Net architectures, and train them on genuine geospatial problems. The course also covers a Siamese U-Net — a specialized architecture designed specifically for change detection that processes two images simultaneously.A key methodological focus throughout is doing things correctly. Every module uses proper spatial train/test splits to prevent data leakage, ensuring models are evaluated on geographically distinct areas they have never seen. This is how professional geospatial deep learning is done in the real world — and it is what separates this course from generic image segmentation tutorials.By the end of this course you will have:Built and trained U-Net models in PyTorch for pixel-wise segmentationProcessed real satellite, aerial, and LiDAR data end to endImplemented spatial train/test splits for honest model evaluationCreated int