REMSA: Foundation Model Selection for Remote Sensing via a Constraint-Aware Agent

Technische Universität Berlin & BIFOLD
REMSA teaser image

REMSA integrates structured RSFM metadata, tool-augmented reasoning, and clarification loops to provide accurate, constraint-aware recommendations for RS foundation model selection.

✨ Key Features

REMSA introduces several capabilities that distinguish it from existing RS retrieval systems, benchmarks, and LLM agents:

🤖 LLM-Driven Model Selection

Automates RS foundation model (RSFM) selection from natural-language queries using a modular agent workflow.

🗂️ Structured RS-FMD Database

Built on the first schema-guided database of 160+ RS foundation models with detailed metadata on modalities, sensors, pretraining, and benchmarks.

🎯 Constraint-Aware Ranking

Combines rule-based filtering with in-context LLM ranking to handle modality, sensor, resolution, compute budget, and performance requirements.

🔍 Clarification & Reasoning Loops

Asks targeted clarification questions for incomplete or ambiguous queries, improving recommendation quality.

🧠 Benchmark-Validated Performance

Evaluated on 100 real RSFM selection scenarios with 900 expert judgments, outperforming retrieval-based and RAG baselines.

Abstract

Foundation Models (FMs) are increasingly used in remote sensing (RS) for applications such as environmental monitoring, land-use mapping, and disaster response. However, selecting an appropriate RS foundation model (RSFM) is challenging due to heterogeneous documentation, diverse modality support, and complex deployment constraints. We introduce RS-FMD, the first structured and schema-guided database of over 160 RS foundation models, covering modality, sensor support, architecture, pretraining datasets, and benchmarks. Built on top of RS-FMD, we propose REMSA, the first LLM-based agent that performs automated RSFM selection from natural-language queries. REMSA interprets user intent, clarifies missing constraints, retrieves candidate models, ranks them via in-context reasoning, and generates transparent explanations. To evaluate REMSA, we construct a benchmark of 100 expert-verified remote-sensing scenarios, generating 900 scored model-query pairs using a novel rubric. Across all metrics, REMSA outperforms baselines including naive agents, dense retrieval, and unstructured RAG systems, demonstrating its utility for real RSFM selection workflows.

📊 REMSA Benchmark Coverage

Category Examples of Scenarios Modalities Typical Constraints
Land Cover Classification Sentinel-1/2, PlanetScope, MODIS, HSI, nighttime imagery SAR, multispectral, hyperspectral, optical Few-shot labels, CPU-only, limited compute, OA-focused
Segmentation & Mapping Flood mapping, wildfire detection, high-res urban expansion, roads SAR, multispectral, thermal, stereo High mIoU, fine spatial detail, real-time inference
Change Detection Urban growth, coastal change, wildfire recovery SAR, multispectral Low compute, few-shot labels, embedded devices
Object Detection Buildings, ports, industrial facilities High-res optical High mAP, low false positives
Multi-Modal Fusion LiDAR + optical, SAR + optical, image + text LiDAR, optical, SAR, text High accuracy, robust fusion required
Self-Supervised / Unsupervised Unlabeled SAR, HSI, LiDAR (Amazon/Arctic/Europe) SAR, hyperspectral, LiDAR SSL compatibility, strong representations
Zero-Shot / Out-of-the-Box Flood mapping, deforestation, canopy height, HSI crops SAR, multispectral, LiDAR, text-image No labels available
Resource-Constrained Inference Jetson Nano, Raspberry Pi, laptop-only workflows SAR, multispectral Lightweight models, fast inference

📈 Quantitative Results

System Avg Top-1 ↑ HQ Hit ↑ MRR ↑
REMSA (Ours) 75.76 40.00% 0.34
REMSA-Naive 72.67 37.33% 0.29
DB-Retrieval 67.37 17.33% 0.23
Unstructured-RAG 71.23 30.67% 0.24

BibTeX

@misc{chen2026remsafoundationmodelselection,
      title={REMSA: Foundation Model Selection for Remote Sensing via a Constraint-Aware Agent}, 
      author={Binger Chen and Tacettin Emre Bök and Behnood Rasti and Volker Markl and Begüm Demir},
      year={2026},
      eprint={2511.17442},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.17442}, 
}