Add FocalLoss, class weights, and enhanced tile filtering

[giswqs]

Summary

Implements the enhancements from #335 to support training with discrete class landcover data.

Changes

New in train.py:

• FocalLoss class — focal loss for class imbalance (FL = -α(1-pt)^γ log(pt)), with configurable alpha, gamma, ignore_index (int or False to disable), reduction, and per-class weights. Uses F.log_softmax + F.nll_loss internally.
• get_loss_function() — factory returning configured CrossEntropyLoss or FocalLoss with flexible ignore_index and optional class weights.
• compute_class_weights() — computes per-class weights from label tiles with inverse-frequency mode, custom multipliers, and weight capping.
• train_segmentation_model() — wired with new parameters: loss_function, ignore_index, use_class_weights, class_weights, custom_class_multipliers, max_class_weight, use_inverse_frequency, focal_alpha, focal_gamma.

New in utils.py:

• min_feature_ratio parameter on export_geotiff_tiles() — filters out tiles where the ratio of non-background pixels is below a threshold (only applies when skip_empty_tiles=True and label data is provided). Summary stats are reported at the end.

New in __init__.py:

• Exports FocalLoss, get_loss_function, compute_class_weights from train.py
• Exports LandcoverCrossEntropyLoss, landcover_iou, get_landcover_loss_function, train_segmentation_landcover from landcover_train.py
• Exports export_landcover_tiles from landcover_utils.py

Backward Compatibility

All new parameters have sensible defaults — existing code continues to work without changes.

Closes #335

---

Closes #335

[github-actions]

🚀 Deployed on https://697c96880a5be10bbd1b5820--opengeos.netlify.app

[Copilot]

Pull request overview

This PR extends the training and tiling utilities to better support highly imbalanced, discrete landcover segmentation workflows, and exposes the new landcover-specific helpers at the package root.

Changes:

• Adds a general-purpose FocalLoss, get_loss_function, and compute_class_weights in train.py, and wires them into train_segmentation_model to support focal loss, class weighting, and flexible ignore_index.
• Extends export_geotiff_tiles in utils.py with a min_feature_ratio parameter to skip tiles with too few non-background pixels and report ratio-based skipping statistics.
• Updates __init__.py to export the new training helpers and landcover-specific utilities (LandcoverCrossEntropyLoss, landcover_iou, get_landcover_loss_function, train_segmentation_landcover, export_landcover_tiles).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
geoai/train.py	Introduces FocalLoss, a configurable loss factory get_loss_function, class-weight computation via compute_class_weights, and extends train_segmentation_model to use these for class-imbalanced landcover training.
geoai/utils.py	Adds min_feature_ratio validation, per-tile ratio filtering, and summary stats reporting to export_geotiff_tiles to drop mostly-background tiles.
geoai/__init__.py	Re-exports the new loss/weight utilities from train.py and landcover-specific training and tiling helpers for easier package-level access.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The function signature declares a return type of torch.nn.Module, but the implementation does not return a model object (it relies on side effects and saving to output_dir), and the docstring states that it returns None. Please align the return type annotation and documentation with the actual behavior (either return the model or change the annotation/docstring to indicate that the function returns None).

[Copilot]

When in_class_data is a raster, stats['feature_pixels'] is incremented as soon as a tile has any features, but tiles can still be skipped later by the min_feature_ratio filter. As a result, the "Average feature pixels per tile" summary can include pixels from tiles that were not actually exported, which makes this statistic misleading; it would be more accurate to count feature pixels only for tiles that are written.

Suggested change

	print(
	f"Average feature pixels per tile: {stats['feature_pixels']/stats['tiles_with_features']:.1f}"
	)
	if min_feature_ratio is False:
	print(
	f"Average feature pixels per tile: {stats['feature_pixels']/stats['tiles_with_features']:.1f}"
	)
	else:
	print(
	"Average feature pixels per tile is not reported when a min_feature_ratio filter is applied."
	)

[Copilot]

compute_class_weights uses Counter here but Counter is never imported in this module, so calling this function will raise a NameError at runtime. Please import Counter from the collections module (or remove the annotation and use a plain dict) so that this function works correctly.

[gw0ods]

Hello Professor Wu, I tried reinstalling geoai on another system and the modules I added didnt seem to work properly... I had to change a few things in them. Sorry about that, I cant push a working version next week I think.
I also added a better IOU method for this discrete landcover classification. The issue was that when there is sparse data false positives shouldnt be considered in IOU calculation and model selection.

[giswqs]

@gw0ods Can you open a new issue and describe what is not working?