GeoCrossBench: Cross-Band Generalization for Remote Sensing
Hakob Tamazyan, Ani Vanyan, Alvard Barseghyan, Anna Khosrovyan, Evan Shelhamer, Hrant Khachatrian

TL;DR
This paper introduces GeoCrossBench, a new benchmark for evaluating remote sensing models' ability to generalize across different satellites and spectral bands, and proposes ChiViT, a self-supervised model extension that improves cross-satellite performance.
Contribution
It presents GeoCrossBench for comprehensive cross-satellite evaluation and develops ChiViT, a self-supervised model that enhances generalization to new satellites and spectral bands.
Findings
Existing foundation models do not outperform general-purpose models in in-distribution settings.
Models experience 2-4x performance drop when generalizing to satellites with no band overlap.
Fine-tuning the last linear layer with oracle labels improves cross-satellite performance.
Abstract
The number and diversity of remote sensing satellites grows over time, while the vast majority of labeled data comes from older satellites. As the foundation models for Earth observation scale up, the cost of (re-)training to support new satellites grows too, so the generalization capabilities of the models towards new satellites become increasingly important. In this work we introduce GeoCrossBench, an extension of the popular GeoBench benchmark with a new evaluation protocol: it tests the in-distribution performance; generalization to satellites with no band overlap; and generalization to satellites with additional bands with respect to the training set. We also develop a self-supervised extension of ChannelViT, ChiViT, to improve its cross-satellite performance. First, we show that even the best foundation models for remote sensing (DOFA, TerraFM) do not outperform general purpose…
Peer Reviews
Decision·Submitted to ICLR 2026
The paper is written in plain English, with figures and tables underlining the findings from the experiments conducted. The appendix provides additional details supporting reproduction of the work.
# Soundness - 2: fair # Presentation - 3: good # Contribution - 2: fair # Strengths The paper is written in plain English, with figures and tables underlining the findings from the experiments conducted. The appendix provides additional details supporting reproduction of the work. # Weaknesses I appreciate the author's effort towards benchmarking the utility of GFMs. However, the current manuscript resembles just a slight variation of work that has been previously published under the same Ge
1. The paper is well written and interesting 2. The considered problem (cross-satellite generalization for remote sensing) is relevant as Earth observation systems diversify rapidly. 3. The proposed self-supervised, band-sampling extension to ChannelViT is an interesting approach to improve transferability.
1. Improvements over DinoV3 are not significant 2. Metrics between methods show an important variance, making the results difficult to read
- Generalization across diverge satellite image sensors is a very important problem that could unlock the use of diverse data and improve the use of newly launched satellites from the get go, this paper advances research towards this important direction by releasing a dataset and proposing an evaluation framework. - The evaluation framework presented is sensible and covers train-test discrepancy by separating cases to in-distribution, no-overlap bands and superset of bands which makes sense from
- The limitation of >100M parameter models seems quite restrictive for the task at hand since it excludes the best performing foundation models. - The proposed evaluation tasks are limited but can be expanded in followup works. - The value of this work depends greatly on the quality of the proposed dataset, it would be nice to see some samples as part of this submission but none were included. - Figure captions are not sufficient for understanding the figures, they should be expanded to include
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSatellite Communication Systems · Remote-Sensing Image Classification · Precipitation Measurement and Analysis
