Benchmarking Geospatial Foundation Models Against a Supervised Baseline for Fine-Grained Alpine Habitat Change Detection with Very High-Resolution Imagery and LiDAR

Geospatial foundation models (GFMs) are proposed for operational Earth observation but benchmarked mostly on urban and agricultural data, so their value and limitations for fine-grained ecological monitoring, with fuzzy class boundaries and severe class imbalance, are largely untested. We benchmark five GFMs (Clay v1.0, Clay v1.5, Prithvi-EO-2.0, DOFA, TerraMind) against two supervised U-Net baselines, one with a transformer (MiT-B2) and one with a convolutional (ResNet-50) encoder, for post-classification habitat change detection in Gesäuse National Park, Austria. All models are trained with three seeds on 2013 expert-interpreted labels (23 habitat classes), using very high-resolution red, green, blue (RGB) and near-infrared aerial imagery and a normalised digital surface model (nDSM) derived from airborne light detection and ranging (LiDAR). Applied to 2019 imagery and 2020 LiDAR, their predictions are compared with the 2013 reference to yield change maps with eight transition classes plus no change. All five GFMs are robustly ahead of the convolutional U-Net, but so is the transformer U-Net, which only Clay v1.0 exceeds, by a small margin (multi-class change macro Intersection over Union (IoU) 0.233 vs. 0.209) whose paired 95% confidence interval (+0.001 to +0.045) nearly includes zero. Adding the nDSM helps every model, but mostly in the canopy-cover classes, whose reference labels are themselves LiDAR-derived. Rare or heterogeneous transitions remain poorly detected by every model. Rule-based post-processing defined with the Gesäuse National Park administration converts the predictions into polygon-structured, ecologically plausible change maps that meet the park’s data standards, with little effect on pixel accuracy (multi-class macro IoU −0.03 to +0.04) but lower recall for forest density loss and setback; the maps are therefore suited to flagging candidate changes for expert verification rather than to unsupervised use.

Authors

Institutions

Publication Details

Journal
Remote Sensing
Published
2026-11-19
DOI
https://doi.org/10.3390/rs18223975
Primary Topic
Remote-Sensing Image Classification
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Benchmarking Geospatial Foundation Models Against a Supervised Baseline for Fine-Grained Alpine Habitat Change Detection with Very High-Resolution Imagery and LiDAR

Manuela Hirschmugl, Harald Kristen, Daniel Kulmer
Remote Sensing
Remote-Sensing Image Classification
article

Benchmarking Geospatial Foundation Models Against a Supervised Baseline for Fine-Grained Alpine Habitat Change Detection with Very High-Resolution Imagery and LiDAR

Manuela Hirschmugl, Harald Kristen, Daniel Kulmer
article en

Abstract

Geospatial foundation models (GFMs) are proposed for operational Earth observation but benchmarked mostly on urban and agricultural data, so their value and limitations for fine-grained ecological monitoring, with fuzzy class boundaries and severe class imbalance, are largely untested. We benchmark five GFMs (Clay v1.0, Clay v1.5, Prithvi-EO-2.0, DOFA, TerraMind) against two supervised U-Net baselines, one with a transformer (MiT-B2) and one with a convolutional (ResNet-50) encoder, for post-classification habitat change detection in Gesäuse National Park, Austria. All models are trained with three seeds on 2013 expert-interpreted labels (23 habitat classes), using very high-resolution red, green, blue (RGB) and near-infrared aerial imagery and a normalised digital surface model (nDSM) derived from airborne light detection and ranging (LiDAR). Applied to 2019 imagery and 2020 LiDAR, their predictions are compared with the 2013 reference to yield change maps with eight transition classes plus no change. All five GFMs are robustly ahead of the convolutional U-Net, but so is the transformer U-Net, which only Clay v1.0 exceeds, by a small margin (multi-class change macro Intersection over Union (IoU) 0.233 vs. 0.209) whose paired 95% confidence interval (+0.001 to +0.045) nearly includes zero. Adding the nDSM helps every model, but mostly in the canopy-cover classes, whose reference labels are themselves LiDAR-derived. Rare or heterogeneous transitions remain poorly detected by every model. Rule-based post-processing defined with the Gesäuse National Park administration converts the predictions into polygon-structured, ecologically plausible change maps that meet the park’s data standards, with little effect on pixel accuracy (multi-class macro IoU −0.03 to +0.04) but lower recall for forest density loss and setback; the maps are therefore suited to flagging candidate changes for expert verification rather than to unsupervised use.

Remote SensingVol. 18(22)
Joanneum Research (AT), University of Graz (AT)
Remote-Sensing Image Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon