🛰️ Remote Sensing

🧪 ICML2026 · 3 paper notes

📌 Same area in other venues: 📷 CVPR2026 (63) · 🔬 ICLR2026 (11) · 🤖 AAAI2026 (7) · 🧠 NeurIPS2025 (12) · 📹 ICCV2025 (11) · 🧪 ICML2025 (7)

Any2Any: Unified Arbitrary Modality Translation for Remote Sensing

Any2Any transforms remote sensing (RS) translation between RGB, SAR, NIR, MS, and PAN from a collection of paired models into a unified latent diffusion model within a shared latent space. By utilizing the million-level RST-1M dataset and target modality residual adapters, it achieves superior fidelity and generalization across 14 seen translation directions and multiple unseen modality combinations.

Localized, High-resolution Geographic Representations with Slepian Functions

This paper constructs a geographic positional encoder that concentrates representation capacity on a Region of Interest (ROI) using spherical Slepian functions. It proposes a Slepian-Spherical Harmonic (SH) hybrid encoding to simultaneously capture local high-resolution details and global coarse-grained context. It consistently outperforms mainstream baselines such as SH, Wavelets, and RFF across five classification, regression, and image-enhancement prediction tasks.

The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench

The authors observe that Vision Foundation Models (VFMs) "appear" to predict satellite imagery well but collapse along physical axes in extreme regimes. By formalizing "scientific alignment" as "structural isomorphism," they release TC-Bench—a global tropical cyclone benchmark—and a three-tier linear probing suite (Static/Dynamic/Constraint) to reveal representation collapse in frozen backbones like DINO, CLIP, SigLIP, and MAE for intense cyclones where \(P_c<980\) hPa.