TerraHeight-S

TerraHeight-S is a compact remote-sensing height estimation model for predicting dense per-pixel Above Ground Level (AGL) height maps from overhead RGB imagery.

The model was created by fine-tuning the pretrained Depth Anything V2 Small (ViT-S) model exclusively on the GAMUS dataset.

TerraHeight-S contains only 24.79M parameters while achieving performance competitive with substantially larger state-of-the-art height estimation models.

Model Details

Property	TerraHeight-S
Task	Single-view remote-sensing height estimation
Input	Overhead RGB imagery
Output	Dense height map
Output units	Metres Above Ground Level (AGL)
Training dataset	GAMUS
Backbone	Depth Anything V2 Small
Encoder	ViT-S
Parameters	24,785,089
Feature dimension	64
Decoder channels	48, 96, 192, 384
Training crop size	630 × 630
Best checkpoint epoch	30
Test-time augmentation	None

Validation Performance

The released checkpoint achieved the following results on the held-out GAMUS validation data.

Height Range	MAE ↓	RMSE ↓	Pearson Correlation ↑
All pixels	1.312 m	2.616 m	0.9238
> 1 m	2.421 m	3.569 m	0.8923
> 5 m	2.693 m	3.951 m	0.8758

The overall evaluation covered approximately 900.7 million valid pixels.

Comparison with Depth2Elevation

Depth2Elevation is a published state-of-the-art reference for single-view remote-sensing height estimation on GAMUS.

Using the GAMUS test benchmark, the following results provide a useful comparison:

Model	Parameters	MAE ↓	RMSE ↓
Depth2Elevation	99.54M	1.991 m	3.489 m
TerraHeight-S	24.79M	1.580 m	3.725 m

TerraHeight-S achieves a lower MAE while remaining close in RMSE, despite using approximately 4× fewer parameters.

This represents roughly a 75% reduction in parameter count compared with Depth2Elevation.

Architecture

TerraHeight-S is based on the pretrained Depth Anything V2 Small architecture.

Rather than training a vision transformer from scratch, the pretrained representation learned by Depth Anything V2 is fine-tuned for remote-sensing metric height estimation using GAMUS RGB imagery and corresponding AGL targets.

The released model configuration is:

Encoder:      ViT-S
Features:     64
Out channels: [48, 96, 192, 384]
Parameters:   24.79M