botao0818/zasolar-exp003-c-targeted-hn

ZAsolar Exp003-C Targeted Hard Negatives

Mask R-CNN checkpoint for rooftop solar installation footprint segmentation in South Africa.

This model is the Exp003-C variant from the ZAsolar project. It was trained with targeted hard negatives extracted from reviewed false-positive detections, following the installation-level footprint segmentation task definition used in this repository.

Files

• best_model.pth: trained PyTorch checkpoint

Training Summary

• Task: installation-level rooftop solar footprint segmentation
• Base architecture: Mask R-CNN
• Training variant: targeted hard negatives (exp003_C_targeted_hn)
• Training data:
  • 4,740 positive chips
  • 883 targeted hard-negative chips
  • 7,720 annotations
• Validation set:
  • 2,640 images
  • 2,948 annotations

Experiment Notes

According to the project experiment log, this targeted hard-negative variant substantially reduced false positives in independent inference evaluation compared with the earlier baseline, improving fixed-threshold precision and F1 while keeping recall reasonably close.

Repository reference:

• Experiment doc: docs/experiments/exp_003_hard_negatives.md
• Training entrypoint: train.py

Intended Use

This checkpoint is intended for research and internal geospatial analysis workflows around rooftop solar detection. It is not packaged as a drop-in Transformers model and should be loaded by the project's PyTorch inference/training code.

Limitations

• Trained for the ZAsolar data and labeling conventions
• Evaluation conclusions depend on the repository's explicit installation profile
• Generalization outside the source geography, imagery characteristics, and annotation policy is not guaranteed