ML Engineer specializing in geospatial computer vision for satellite imagery. End-to-end pipelines: raw Sentinel-2/HLS → field boundary segmentation → crop classification → GeoPackage.
| What | Numbers |
|---|---|
| Field segmentation model (ResUNet-A) | MCC = 0.91, 5+ countries |
| Production pipeline (Airflow + AWS) | 35,000+ EOPatches, ~6M km², 11 Brazilian states |
| Transfer learning: global → local | Brazil MCC 0.77 → 0.89, Argentina MCC 0.87 |
| Postprocessing optimization | 33h → 18h (memray, malloc_trim, multiprocessing) |
| Inference optimization (ONNX) | TF→ONNX 35-40% speedup; TFLite quantization research |
| Crop classification (Transformer Encoder) | safra/safrinha, validated vs CONAB/SIDRA |
Python PyTorch TensorFlow ONNX TFLite Apache Airflow AWS (EC2 · S3 · Batch)
GDAL Rasterio GeoPandas eo-learn STAC Docker Sentinel-2
satellite-field-segmentation — Field boundary detection from satellite imagery: ResUNet-A, geospatial postprocessing, GeoPackage output.