Research map / Vision-language and multimodal / Multimodal LLMs

TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TerraScope, an InternVL3-based VLM with a SAM-2 mask decoder that interleaves segmentation masks with reasoning for pixel-grounded earth observation reasoning, trained on the 1M-sample Terra-CoT dataset and evaluated on the new TerraScope-Bench.

Problem

Vision-language models struggle with fine-grained, pixel-accurate geospatial reasoning over earth observation imagery, including multi-modal (optical/SAR) and multi-temporal analysis, and existing EO benchmarks do not assess pixel-level reasoning.

Approach

Generate segmentation masks via a [SEG]-triggered mask decoder during reasoning, then inject mask-selected visual tokens back into the LLM's reasoning sequence ("thinking with pixels"), with text-guided selection between optical and SAR tokens and explicit timestamps for multi-temporal inputs.

Key results

• TerraScope reaches 68.9 average on TerraScope-Bench versus 43.3 for the best general VLM (Qwen3-VL-Think) and 59.6 for fine-tuned GLM-4.1V-Think.
• Correct predictions have mean mask IoU 0.628 versus 0.443 for incorrect ones (Pearson r = 0.607).
• Ablation: Box CoT reaches 62.8 vs 68.9 for pixel-mask TerraScope on TerraScope-Bench; text-guided modality selection improves DM to 68.2 vs 65.9 with no fusion.

Surprising findings

• Enabling text-guided modality selection only at test time yields no improvement, so training with it is required.
• EO-specific VLMs do not significantly outperform general VLMs on TerraScope-Bench.

Limitations

• Interleaved mask and reasoning generation increases context length during training and inference.
• Handles only optical RGB and SAR modalities, not multispectral or hyperspectral data.
• Temporal reasoning is limited to bi-temporal analysis.

Experimental setup

Datasets: BigEarthNet ChatEarthNet xBD TEOChat RSVQA-LR Landsat30-AU DisasterM3

Benchmarks: TerraScope-Bench Landsat30-AU DisasterM3

Metrics: accuracy mean IoU Pearson correlation

Baselines: GPT-4o Qwen3-VL-Think Qwen2.5-VL InternVL3 GLM-4.1V-Think LLaVA-OV GeoChat EarthDial EarthMind Mint-CoT ThinkGeo EarthAgent

Builds on: InternVL3 SAM-2 LISA Mint-CoT GRIT

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.91)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.90)
• GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.89)
• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.89)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.89)
• TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data (ICLR 2025 · similarity 0.89)