chrieke/InstanceSegmentation_Sentinel2

Implements a fully convolutional instance segmentation architecture (FCIS) on MXNet trained with Sentinel-2 multispectral imagery and COCO-format annotations from Denmark's LPIS field database. Processes raw RGB bands without extensive preprocessing, achieving generalization across varying field geometries while supporting multi-task learning for simultaneous crop-type classification. Includes preprocessing pipelines for satellite data ingestion, configurable ResNet-101 backbone training, and evaluation workflows compatible with AWS GPU instances.

430 stars. No commits in the last 6 months.