Random forests, Segment Anything and U-Nets on satellite imagery: training data, spatially honest validation and inference on full scenes.
Map blanket bog, heath, bracken, grassland, woodland, water and built-up land in the Peak District from 2023 Sentinel-2 seasonal composites with an eight-rule spectral index classifier, then assess it properly: a confusion matrix, user's and producer's accuracy, kappa and its caveats, block-bootstrap confidence intervals, and why a random train/test split overstates accuracy.
Classify nine upland habitats in the Peak District with LightGBM on 77 multi-temporal features: seasonal Sentinel-2 bands and indices, Sentinel-1 VV and VH backscatter and Copernicus DEM terrain. Tune with spatial-fold cross-validation and class weights, run an S2 / S2+S1 / S2+S1+terrain ablation, explain the model with TreeSHAP and permutation importance, and compare it with a random forest and the Part 1 rules on the same test blocks.
Use Google DeepMind's AlphaEarth Foundations satellite embeddings, a geospatial foundation model, in Python without a login. Classify nine Peak District habitats with a linear model, compare a few-shot learning curve against raw Sentinel-2 bands, map uncertainty, check calibration, simulate active learning, and estimate habitat areas with confidence intervals from a stratified random sample (Olofsson et al. 2014).
Segment 2023 Sentinel-2 seasonal composites of the Peak District into image objects with scikit-image SLIC and Felzenszwalb, tune the segment size against Living England boundaries, compute per-object statistics, shape and GLCM texture with NumPy, classify the objects with LightGBM and compare OBIA with a pixel-based classifier on accuracy, salt-and-pepper noise and patch counts. Then export the habitat objects as GeoPackage polygons.
Classify habitats from a full year of cloudy, irregular Sentinel-2 observations with two temporal neural networks in PyTorch: a 1D convolutional TempCNN and a lightweight temporal attention encoder (L-TAE) with day-of-year positional encoding. Compare masking with gap-filling, test against a LightGBM baseline on spatial blocks, read which dates matter from attention and occlusion, and predict a wall-to-wall 10 m map on a CPU.
Run a trained PyTorch U-Net over a whole Sentinel-2 scene of a city it has never seen: tiled sliding-window inference on a large GeoTIFF with overlapping, blended windows, a georeferenced Cloud Optimised GeoTIFF of classes and confidence, raster-to-vector polygons with rasterio.features.shapes in a GeoPackage, and an honest check against reference labels.
Train a U-Net to segment Sentinel-2 imagery into buildings, built-up land, trees, grass and water with PyTorch, segmentation_models_pytorch and TorchGeo's Sentinel-2 pretrained ResNet-18 weights, on a CPU in five minutes. Real loss curves, per-class IoU and F1, a confusion matrix and an honest look at what limits accuracy.
Build a labelled deep learning dataset from free data: stream a Sentinel-2 L2A composite from the Planetary Computer, rasterise Overture building footprints over ESA WorldCover, cut georeferenced image chips, split them by spatial block so the test set does not leak, check the class balance and load it all as a TorchGeo dataset.
Run Meta's Segment Anything Model on high-resolution aerial imagery with segment-geospatial (samgeo), on a CPU. Generate masks automatically, vectorise them to a GeoPackage, filter them by area, shape and NDVI to find buildings, and prompt SAM with points and boxes, checked against OpenStreetMap.
Supervised land cover classification in Python: build seasonal Sentinel-2 composites with odc-stac, sample labels from ESA WorldCover, train a scikit-learn random forest, test it with spatial cross-validation and map Sheffield and the Peak District.