Research Project · Farmland Remote Sensing

FarmSeeker

Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain

Haiyang Wu^1,*, Weiliang Mu^1,*, Zhuofei Du¹, Dandan Zhong¹, Kaijie Shi¹, Haifeng Li¹, Chao Tao^1,†

¹ Central South University, Changsha, China

ocean-w@csu.edu.cn, 245012170@csu.edu.cn, kingtaochao@csu.edu.cn

^* Equal contribution ^† Corresponding author

Code Paper · Coming Soon Dataset · Coming Soon Video Demo

Official project page. Paper, checkpoints, and benchmark links will be activated upon release.

Video Demo

FarmSeeker Video Demo

docs/assets/videos/FarmSeeker_Demo.mp4

Current image → initial mask → ambiguity perception → evidence query → collaborative reasoning → refined mask

Motivation

Is segmentation failure really just because the model does not see well enough?

The answer may not be in the current image at all. A natural solution is to seek extra information gain-not by looking at more images indiscriminately, but by seeing the most critical evidence when it is truly needed.

Limited by observation time and spatial field of view, a single image may not contain the complete evidence required for reliable farmland recognition. FarmSeeker neither repeatedly mines the current image nor blindly introduces more inputs. Instead, it first localizes segmentation ambiguities, determines whether temporal cues or broader spatial context are missing, and then acquires on demand the extra spatio-temporal evidence expected to provide the greatest information gain. The newly acquired evidence is ultimately translated into more reliable farmland segmentation results.

Method

FarmSeeker answers when extra evidence is needed, what evidence to retrieve, how to acquire it, and how to use it.

Overview of FarmSeeker. The framework performs initial segmentation, ambiguity perception, evidence requirement decision, image retrieval, collaborative reasoning, and local mask refinement in a closed loop.

Initial perception

Generate an initial farmland mask from the current remote-sensing image.

Ambiguity perception

Locate regions where the current observation may not support a reliable decision.

Evidence requirement

Decide whether temporal, spatial, joint, or no additional evidence is required.

Image retrieval

Query aligned multi-temporal imagery or a wider spatial-context observation.

Collaborative refinement

Reason across current and retrieved images, then locally correct the mask with SAM 2.1.

Benchmark

We build GSFS-Bench to evaluate ambiguity perception, evidence selection, and collaborative reasoning under realistic farmland ambiguity.

Scale

10K

Evaluation samples cropped to 512×512 from high-resolution remote-sensing imagery.

Resolution

0.5–1 m

High-resolution observations suitable for fine-grained farmland interpretation.

Coverage

Global

Scenes span multiple major agricultural regions with diverse climates and field patterns.

Evidence Pool

Each sample is associated with spatial and temporal evidence pools for on-demand retrieval.

Illustration of GSFS-Bench, including evaluation samples, ambiguity annotations, and spatio-temporal evidence pools.

Subset	Purpose	Annotation / Supervision
General Evaluation Set	Farmland segmentation performance	Pixel-level farmland masks
Ambiguity Perception Subset	Evaluate whether the model can localize unreliable regions	Reference ambiguity regions
Tool-Calling Subset	Evaluate whether the model requests the correct evidence type	Temporal / spatial / none labels
Collaborative Reasoning Subset	Evaluate whether the model can use retrieved images to revise its decision	Multi-image reasoning samples

Results

Qualitative comparisons demonstrate the effectiveness of FarmSeeker in resolving both spatio-temporal ambiguity and challenging farmland boundaries across diverse agricultural regions.

Qualitative comparison of FarmSeeker and mainstream farmland segmentation methods — **Qualitative comparison of farmland segmentation between FarmSeeker and mainstream methods across different agricultural regions.** (a) Selected segmentation results on the Northeast China Plain. (b) Selected segmentation results on the Huang-Huai-Hai Plain. (c) Selected segmentation results in the Northern Arid and Semi-arid Region. (d) Selected segmentation results on the Loess Plateau. (e) Selected segmentation results in South China Areas. (f) Selected segmentation results in the Sichuan Basin. (g) Selected segmentation results on the Yunnan-Guizhou Plateau. (h) Selected segmentation results on the Middle-Lower Yangtze Plain. Red circles indicate regions with notable differences. Green masks represent farmland, while black denotes the non-farmland background.

Citation

The final BibTeX entry will be updated after the paper record becomes public.

@misc{wu2026thinkextraimagefarmlandsegmentation,
              title={Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain}, 
              author={Haiyang Wu and Weiliang Mu and Zhuofei Du and Dandan Zhong and Kaijie Shi and Haifeng Li and Chao Tao},
              year={2026},
              eprint={2607.28186},
              archivePrefix={arXiv},
              primaryClass={cs.CV},
              url={https://arxiv.org/abs/2607.28186}, 
        }