[test]: Add background class label and train with background supervision

[Dinghye]

We want to explicitly model the background class as an additional supervised signal. Currently, background pixels are implicitly handled, but adding a dedicated background label may help the model better distinguish foreground segmentation targets from non-target regions.

Motivation

The goal is to improve segmentation quality, especially by reducing false positives and helping the model learn clearer decision boundaries between foreground objects and background regions.

Tasks

• Define how to compute the background class from existing labels.
• Add the background class as a new label/channel in the training targets.
• Update the data loading pipeline to include the background label.
• Modify the loss function if needed to supervise the background class.
• Train and compare against the current best baseline.

Deliverables

• Code changes for background label generation.
• Updated training pipeline supporting background supervision.
• Experiment results comparing:
  • current baseline(run in your env)
  • baseline + background supervision
• Short summary of whether this improves IoU / RMSE.

Acceptance Criteria

• Background labels are correctly generated and visualized on sample images.
• Training runs successfully with the new background supervision.
• Results are reported with IoU and RMSE comparison against baseline.

Notes

Please keep experiments reproducible:

• record config changes
• record checkpoint path
• report validation IoU and RMSE
• attach qualitative visualization if possible

[fresleven]

Experiment	iou_bld	iou_tree	iou_wat	RMSE_bH	RMSE_vH	Score
original	0.5116	0.7635	0.4842	1.8802	3.2935	0.4766
bkg (v1)	0.5069	0.7635	0.4975	1.8869	3.3278	0.4755
bkg (v2)	0.5126	0.7645	0.4686	1.8930	3.3363	0.4719

These models are using alphaearth_tessera_iou_fusion_N_base48 (not the newest one).
I tried testing out the performance of using softmax instead of sigmoid on fraction_head. I utilized soft cross entropy loss to deal with the fractions. Overall, it doesn't improve the overall score and actively hurts the height measurement. I have two runs:

• bkg (v1): Just replaces sigmoid with softmax and changes the loss function
• bkg (v2): Also edits the height measurement. Edits p_fg to be 1 - p_bkg from the fraction head.

Reproduciblitiy:

• bkg (v1) can be reproduced with commit b168875
• bkg (v2) can be reproduced with commit dd8595e

Future directions:

• bkg (v2) has the only interesting result of increasing the IoU of buildings and vegetation at the expense of height RMSE. This suggests that creating a height estimation more indicative of true probabilities might help IoU
• These results may be indicative that the current height estimation doesn't take true proportions into account (which it shouldn't it should just take the highest point in that area)

[Dinghye]

@fresleven @weihv How about we just use the background for presence head(add a BCE loss for background) and keep the fraction head for the original 3 channels? Because the fraction head is an aux head, not our target, and will directly affect the height. Do you think it will work?