New benchmark GeoCrossBench targets cross-band generalization for remote sensing models

Researchers have introduced GeoCrossBench, an extension of the GeoBench benchmark designed to evaluate the cross-band generalization capabilities of remote sensing foundation models. This new benchmark includes protocols for standard in-distribution performance, generalization to unseen bands, and generalization to test inputs with a superset of training bands. To support this, a self-supervised model called $\chi$ViT was developed as a baseline. Experiments using 11,900 NVIDIA H100 GPU-hours revealed that while models like DOFA and Vision Transformer Base perform well in specific settings, all models exhibit significant performance degradation when evaluated on unseen bands, highlighting the need for more robust cross-band generalization in future remote sensing models. AI

IMPACT This benchmark and model development could lead to more robust remote sensing AI capable of adapting to new satellite data without costly retraining.

RANK_REASON The cluster describes a new benchmark and a supporting model for remote sensing research, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• ChannelViT
• $\chi$ViT
• DOFA
• GeoBench
• GeoCrossBench
• Hakob Tamazyan
• ImageNet
• NVIDIA H100
• Panopticon
• Vision Transformer Base

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →