Domain Generalization Geospatial Foundation Optical + SAR Segmentation

CrossEarth Series : a project suite for cross-domain geospatial perception

A unified showcase of four research threads—CrossEarth (optical DG), CrossEarth-SAR (SAR DG), CrossEarth-Gate (cross-modal gating), and Earth-Adapter (RS-PEFT)—designed to generalize across regions, sensors, and imaging conditions.

Pillars Papers Contributors

Optical DG evaluation

32 task settings

Across five compositional domain gaps

🛰️

SAR DG evaluation

22 sub-benchmarks

Across eight domain gaps

📡

Quick map

Cross-domain stack

🌍

CrossEarth

TPAMI 2025

Optical RS DG via Earth-Style Injection + Multi-Task Training.

🧩

Earth-Adapter

AAAI 2026

RS-PEFT that mitigates artifacts using frequency-guided MoA.

📡

CrossEarth-SAR

Under Review

SAR DG via physics-guided sparse MoE and large-scale resources.

CrossEarth-Gate

Under Review

Cross-modal gating for unified optical perception.

CrossEarth Series stack

CrossEarth Series spans optical DG, SAR DG, PEFT, and cross-modal fusion. Scroll down to the publications, benchmarks, contributors, and results sections to explore the full research suite.

Congratulations to CrossEarth for being accepted by TPAMI 2025! 2025-12 Congratulations to Earth-Adapter for being accepted by AAAI 2026! 2026-01

✨Why CrossEarth Series

Generalization-first geospatial perception

Remote sensing (optical and SAR) experiences domain shifts from region, sensor, spectral band, platform, polarization, and even climate. CrossEarth Series organizes a research stack that scales from data-centric augmentation, to PEFT for artifacts, to physics-guided sparsity for SAR.

Region Sensor Spectral Climate

Unseen domains Robust representations Earth observation

Design principles

Coverage & diversity
Expand training distributions toward the real-world test manifold.
Lightweight adaptation
Correct domain-specific artifacts without breaking frozen foundation representations.
Physics-aware scaling
Use physically grounded cues to stabilize routing under extreme SAR heterogeneity.

What CrossEarth Series offers

Reproducible Baselines — Standardized codebases and experiment configs for fair comparison.
Modular Components — Pluggable modules (PEFT, gating) that integrate with existing foundation models.
Standardized Evaluation — Unified benchmarks and metrics for optical and SAR domain generalization.

🧭Core pillars

Four pillars, one generalization narrative

Switch between pillars to preview how each work fits into the CrossEarth Series stack.

📚Publications

Four papers that define the CrossEarth Series

Use these cards as your canonical project entry points, with direct links to papers, code, and (where available) project resources.

🌍 Vision Foundation Model

CrossEarth: Geospatial Vision Foundation Model for Domain Generalizable Remote Sensing Semantic Segmentation

TPAMI 2025

Combines Earth-Style Injection and Multi-Task Training to handle diverse unseen domain gaps in optical remote sensing.

✨

✓Earth-Style Injection broadens training-domain coverage via field-style embeddings.
✓Multi-task segmentation + MIM improves robustness under distribution shifts.
✓Curates a comprehensive RSDG segmentation benchmark spanning multiple domain gaps.

RSDGSemantic SegmentationOptical RSMIM

Paper Code

🧩 Parameter-Efficient Fine-Tuning

Earth-Adapter: Bridge the Geospatial Domain Gaps with a Frequency-Guided Mixture of Adapters

AAAI 2026

Tailors PEFT to RS segmentation by isolating and mitigating artifacts using frequency decomposition and dynamic fusion.

✨

✓DFT splits features into LF/HF components to isolate artifact-heavy signals.
✓Mixture of Adapters refines each subspace; router aggregates adaptively.
✓Maintains a strong parameter–performance trade-off for dense prediction.

PEFTAdaptersDFTArtifacts

Paper Code

📡 SAR Foundation Model

CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation

Under Review

Introduces physics-guided sparse MoE routing and large-scale SAR resources to push SAR domain generalization.

✨

✓Physics-guided routing stabilizes expert selection under severe SAR domain shifts.
✓Sparse MoE scales capacity while controlling compute.
✓Builds a large SAR dataset and multi-gap benchmark suite for unified evaluation.

SARSparse MoEPhysicsDG

Paper Code

🔀 Cross-Modal Gating

CrossEarth-Gate: Cross-Modal Gating for Unified Optical Geospatial Perception

Under Review

Introduces a gating module for optical representations, enabling robust cross-modal geospatial understanding.

✨

✓Aligns optical features with lightweight cross-modal gates.
✓Improves transfer across sensors and acquisition conditions when used with CrossEarth.
✓Designed as a reusable module within the broader CrossEarth Series stack.

Cross-ModalGatingOpticalDG

Paper Code

📊Evaluation

Benchmarks built for real-world shifts

Domain gaps are compositional. The CrossEarth Series stack emphasizes standardized, multi-gap evaluation across both optical and SAR modalities.

🛰️

Optical RSDG suite

Curated task settings for semantic segmentation that cover multiple application scenes and compositional domain gaps.

Task settings

Standardized RSDG evaluation for segmentation

Domain gaps

Region / Spectral band / Platform / Climate (+ compositions)

Unseen Region Unseen Spectral Band Unseen Region + Spectral Band Unseen Region + Platform Unseen Region + Climate

📡

SAR DG suite

A unified standard for SAR domain generalization—built to reflect fragmentation across sensors, bands, and polarization.

dataset

CrossEarth-SAR-200K

Large-scale SAR resource for continued pre-training

View on HuggingFace

Sub-benchmarks

Domain gaps

📈

State-of-the-Art performance

+10% mIoU

CrossEarth-SAR achieves leading results across 22 sub-benchmarks and 8 domain gaps, with physics-guided sparse MoE scaling capacity while keeping compute practical.

🖼Results

Visualizing Generalization

Qualitative results across diverse benchmarks: from optical cityscapes to complex SAR geometries.

CrossEarth optical DG qualitative result

Robust Optical Transfer

Achieves robust segmentation on unseen cities (e.g., Dubai, Zurich) by leveraging Earth-Style Injection to bridge regional domain gaps.

OpticalCross-RegionStyle Injection

Earth-Adapter artifact mitigation result

Artifact-Free Adaptation

Mitigates domain-specific artifacts (e.g., sensor noise, spectral shifts) using frequency-guided sparse adapters, outperforming full fine-tuning.

PEFTArtifact RemovalFrequency Domain

Physics-Aware SAR Perception

Overcomes extreme SAR geometry distortions and speckle noise via physics-guided routing, enabling clear semantic interpretation in all-weather conditions.

SARPhysics-GuidedSparse MoE