Get Your Embedding Space in Order: Domain-Adaptive Regression for Forest Monitoring

Sizhuo Li ^* , Dimitri Gominski ^* , Martin Brandt , Xiaoye Tong , Philippe Ciais

University of Copenhagen, LSCE Paris
ECCV 2024
^* Indicates Equal Contribution

DRIFT dataset: The DRIFT dataset includes patch-level tree counts, mean height and tree cover in 5 countries. We hypothesize that ordered relationships generalize better across domains.

Abstract

Image-level regression is an important task in Earth observation, where visual domain and label shifts are a core challenge hampering generalization. However, cross-domain regression with remote sensing data remains understudied due to the absence of suited datasets. We introduce a new dataset with aerial and satellite imagery in five countries with three forest-related regression tasks. To match real-world applicative interests, we compare methods through a restrictive setup where no prior on the target domain is available during training, and models are adapted with limited information during testing. Building on the assumption that ordered relationships generalize better, we propose manifold diffusion for regression as a strong baseline for transduction in low-data regimes. Our comparison highlights the comparative advantages of inductive and transductive methods in cross-domain regression.

Ordered embeddings retain their structure when changing domains.

A. We apply Geometric Order Learning (Lee et al. NeurIPS 2022) on the source domain to build a manifold that evolves with the target value.
B. On the target domain, embeddings become noisier, but retain their order. A few support examples give meaning to the manifold.
C. We apply the proposed MDR (Manifold Diffusion for Regression) to assign values to all data points, in a transductive manner.

Some visual examples from DRIFT. First row: mean non-zero height (m). Second row: tree counts. Third row: tree cover.

DRIFT Dataset

The DRIFT dataset is a large-scale dataset for vegetation monitoring. It includes three variables of interest (tree counts, tree cover, mean non-zero height) in five countries. Aerial images were collected from public national campaigns, the satellite imagery in Slovenia was purchased from Planet's SkySat archive.

Download link

See below for image credits and links to the providers:

Denmark
20cm ground resolution RGB aerial imagery from the Danish Agency for Data Supply and Infrastructure (SDFI), 2018. Refer to the Terms of Use if you wish to reuse the data. CHMs are from the Danmarks Højdemodel product by SDFI, CC BY 4.0 license.
France
20cm ground resolution RGB aerial imagery from the French Institute of Geography (IGN), BD ORTHO© product, 2018-2020. CHMs derived from the LiDAR HD product. All data is under the 2.0 Open License. Refer to the data catalogue for more information.
Slovakia
20cm ground resolution RGB aerial imagery from GKÚ Bratislava, NLC, 2021, downloaded from the GeoPortal. Refer to the Terms of Use if you wish to reuse the data. CHMs derived from the national ALS campaign, credits to ÚGKK SR, CC BY 4.0 license.
Spain
25cm ground resolution RGB aerial imagery and CHMs derived from the Spanish Institute of Geography (IGN), PNOA 2020, scne.es. Data is under the CC-BY 4.0 License. Refer to the Terms of Use if you wish to reuse the data.
Slovenia
50cm ground resolution RGB satellite imagery bought from Planet, SkySat sensor, 2021. Any usage must be solely for Noncommercial education or scientific research purposes, and publication in academic or scientific research journals. All such publications must include an attribution that clearly and conspicuously identifies Planet Labs PBC. CHMs derived from the national ALS campaign, open license with attribution: Ministry of the Environment and Spatial Planning, Slovenian Environment Agency (ARSO).

-->

BibTeX

@InProceedings{DRIFT_2024_ECCV,
        author    = {Li, Sizhuo and Gominski, Dimitri and Brandt, Martin and Tong, Xiaoye and Ciais, Philippe},
        title     = {Get Your Embedding Space in Order: Domain-Adaptive Regression for Forest Monitoring},
        booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
        month     = {October},
        year      = {2024},}