PS-GNN: Graph Neural Networks for Persistent Scatterer Selection
Persistent-scatterer selection as node classification on a spatial graph. F1 0.82 vs 0.69 for StaMPS-style selection on hard test scenes, and 0.97 recall on borderline scatterers.
Overview
From crop types and illegal mining to building footprints and plant counts, I design deep-learning systems for satellite, aerial and drone imagery, and I judge them by how they perform in places they have never seen.
What you get
- Segmentation, detection and classification models trained for your imagery and classes
- Benchmarks against simple baselines, so the added complexity is justified
- Spatially honest accuracy reports with uncertainty, not just a headline score
- Inference pipelines from search to finished map, mask or dataset
- Deployment artifacts: ONNX models, FastAPI / WebSocket inference servers, edge-ready exports
Who this helps
Agriculture & food security
Crop-type maps, crop condition and plant counting for smallholder and commercial farms.
Forest & mining monitoring
Detecting illegal mining, deforestation and land-cover change under persistent cloud.
Governments & NGOs
Land-cover, building and change maps that stand up to scrutiny.
Teams with a model that fails in the field
Diagnosing domain shift, label problems and leakage in existing models.
How I approach it
-
01
Start from the decision, not the architecture
Which map, at what resolution, with what error is acceptable, and who will act on it. That defines the classes, the metric and the validation design.
-
02
Audit the data and labels
Check sources, label noise and class balance; mining scars, for example, covered under 4% of the Atewa landscape. Where labels are scarce, bootstrap them from OpenStreetMap, building footprints, ESA WorldCover, Dynamic World, SAM or UAV imagery.
-
03
Baseline first
Train a ResNet-50 or U-Net baseline before reaching for foundation models. A complex model must beat a simple one on the same spatial split.
-
04
Fit the model to the physics
SAR–optical fusion for cloud-prone regions, multi-temporal stacks for crops, EO foundation models (Clay, Prithvi-EO-2.0, DINOv3) where they add value, and imbalance-aware losses such as focal, Tversky and OHEM with hard negative mining.
-
05
Validate spatially
Spatial-block or agro-ecological-zone cross-validation, held-out tiles and field re-validation. Random pixel splits overstate accuracy because neighbouring pixels are nearly identical.
-
06
Deploy and keep watching
Export to ONNX, serve through FastAPI with batch and WebSocket inference, and monitor drift, with Grad-CAM, SHAP and uncertainty estimates to explain predictions.
Evidence from my work
- 82% / 0.79Recall / F1 for illegal-mining detection on held-out tiles
- > 0.70F1 for 10 m crop types under spatial cross-validation
- 95.8%In-domain score of a crop-disease CNN that failed on smallholder farms, diagnosed and fixed
- 119+Architectures registered in pygeovision
The crop-disease model scored 95.8% in its training domain but close to zero on smallholder farms; geographically stratified data diversification and field re-validation resolved it.
Selected work
View case study
pyunwrap: Learned InSAR Phase Unwrapping
A network that predicts the integer 2π cycle count at every pixel, so it can never contradict the measured phase. 99.1% of pixels exact on easy test scenes, 96.6% on moderate.
Atewa Illegal Mining Detector
Dual-stream U-Net + YOLOv8 fusing Sentinel-1 and Sentinel-2, trained on UAV-derived labels. Recall 82%, F1 0.79 on held-out tiles.
Crop-Type Mapping with EO Foundation Models
Clay fine-tuned on multi-temporal Sentinel-2 stacks for 10 m crop types, benchmarked against ResNet-50; F1 above 0.70 under spatial cross-validation.
What sets my work apart
- 1
Spatial cross-validation by default
Every reported score comes from areas the model did not train on.
- 2
I go looking for failure
Domain shift, leakage and label noise are tested for directly, because that is where field deployments break.
- 3
Uncertainty is a feature
Ensemble and Bayesian uncertainty flag the predictions that need a human look.