--- license: cc-by-4.0 task_categories: - other language: - ko - en tags: - GIS - benchmark - geospatial - LLM - agent - QGIS - MCP - spatial-analysis pretty_name: GIS-Bench size_categories: - n<100 --- # GIS-Bench: A Benchmark for Evaluating LLM-based GIS Automation Agents ## Dataset Description **GIS-Bench** is a benchmark framework designed to evaluate the capability of Large Language Model (LLM)-based agents in automating GIS (Geographic Information System) tasks using the Model Context Protocol (MCP). This benchmark was developed as part of a master's thesis research and targets real-world spatial analysis workflows executable in QGIS Desktop via Claude + MCP integration. - **Paper:** *GIS-Bench: LLM 기반 GIS 자동화 에이전트 평가 벤치마크* (KCC 2026, submitted) - **Repository:** [yiseo0/GIS-Bench](https://huggingface.co/datasets/yiseo0/GIS-Bench) - **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) --- ## Benchmark Overview | Item | Details | |------|---------| | Total Tasks | 50 | | Difficulty Levels | 3 (Level 1 / Level 2 / Level 3) | | Task Type | GIS spatial analysis & visualization | | Output Format | HTML (Leaflet.js map), PNG, JPG | | Evaluation Method | VLM-based automated scoring (Claude Sonnet) | | Scoring | 130-point raw → 100-point normalized | | Pass Threshold | 60 points (normalized) | --- ## Evaluated Models | Model | Condition | Pass Rate (Overall) | |-------|-----------|-------------------| | GPT-5.4-mini | Few-shot | 10.0% | | GPT-5.2 | Few-shot | 36.0% | | Claude (Sonnet) | Zero-shot | 60.0% | | Claude (Sonnet) | Few-shot | 78.0% | | Claude (Sonnet) + QGIS MCP | Agentic | **86.0%** | > Statistical significance confirmed via McNemar's test (p < 0.05) --- ## Task Structure Tasks are organized across 3 difficulty levels: | Level | Description | Example | |-------|-------------|---------| | Level 1 | Basic data loading & visualization | Load a shapefile and display on map | | Level 2 | Spatial filtering & styling | Filter buildings by area > 500㎡ with color coding | | Level 3 | Multi-step spatial analysis (MCP-driven, server data access) | Load Landsat NDWI data via MCP and generate a landslide density heatmap using `native:heatmapkerneldensity` | Each task entry includes: - `task_no`: Task ID (1–45) - `prompt`: Natural language instruction (Korean) - `level`: Difficulty level (1, 2, or 3) --- ## Data Sources All spatial datasets used in this benchmark are sourced from Korean public data portals under open licenses. | 데이터명 | 실제 파일 | 형식 | 공간 유형 | 좌표계 | 출처 | 라이선스 | |---------|---------|------|---------|--------|------|---------| | 서울시 CCTV 설치 현황 | `few_shot_test_서울시 안심이 CCTV 연계 현황` | CSV | Point | EPSG:4326 (WGS84) | [서울 열린데이터광장](https://data.seoul.go.kr) | 공공누리 1유형 | | 서울시 고도지구 경계 | `UQ123_용도지구(고도지구)_20250805` | SHP | Polygon | EPSG:5186 | [국토정보플랫폼](https://www.nsdi.go.kr) | 공공누리 1유형 | | 서울시 행정경계 | `서울시_SIG.shp` | SHP | Polygon | EPSG:5186 | [공공데이터포털](https://www.data.go.kr) | 공공누리 1유형 | | 수치표고모델 (DEM) | `(B080)공개DEM_34602_img_2025` | GeoTIFF | Raster | EPSG:5186 | [국토정보플랫폼](https://www.nsdi.go.kr) | 공공누리 1유형 | | 고해상도 위성영상 | `Landsat_위성영상의_정규물지수_2024` | GeoTIFF | Raster | EPSG:5186 | [NASA EarthData](https://earthdata.nasa.gov) | Public Domain | > All datasets are publicly available and free to use with attribution. --- ## Evaluation Pipeline Evaluation is performed using **VLM-based automated scoring** with Claude Sonnet as the judge model. ### Scoring Criteria (130-point raw scale) #### Basic Quality (100 pts) | Criterion | Max Score | Description | |-----------|-----------|-------------| | `exists` | 20 | Meaningful visualization present | | `accuracy` | 30 | Correct region and dataset used | | `requirement` | 30 | Color, filter, buffer conditions met | | `completeness` | 20 | Legend, title, labels complete | #### Spatial Accuracy (30 pts) | Criterion | Max Score | Description | |-----------|-----------|-------------| | `spatial_location` | 15 | Data displayed in correct geographic area | | `geometry_validity` | 5 | No polygon/buffer distortion | | `numeric_match` | 10 | Numeric conditions reflected in output | **Normalization:** `score_normalized = round(total_raw / 130 * 100)` Each task is evaluated **3 times** (N=3) and the average score is used for robustness. See [`evaluation/eval_criteria.md`](evaluation/eval_criteria.md) for the full evaluation prompt and criteria details. --- ## Reference Outputs Since GIS tasks do not always have a single deterministic "ground truth," this benchmark provides **reference outputs** — screenshots generated by the benchmark author under each task condition. These serve as visual references for what a correct result should look like. Reference screenshots are provided in `assets/screenshots/` organized by model and task ID. --- ## MCP Tool Dependency This benchmark was conducted using **QGIS MCP** to connect Claude to QGIS Desktop. - **Tool:** [jjsantos01/qgis_mcp](https://github.com/jjsantos01/qgis_mcp) - **Description:** MCP server + QGIS plugin that allows LLMs to control QGIS Desktop - **Note:** The MCP tool code is not included in this repository. Please refer to the original repository for installation. --- ## Repository Structure ``` GIS-Bench/ ├── README.md # This file (Dataset Card) ├── data/ │ └── tasks.json # 50 task prompts with level labels ├── evaluation/ │ ├── run_evaluation_vlm.py # VLM evaluation pipeline (5 models) │ └── eval_criteria.md # Scoring criteria & judge prompt └── assets/ └── screenshots/ # Reference output screenshots ``` --- ## Citation If you use GIS-Bench in your research, please cite: ```bibtex @inproceedings{gisbench2026, title = {GIS-Bench: LLM 기반 GIS 자동화 에이전트 평가 벤치마크}, author = {김이소 and 장두성}, booktitle = {한국정보과학회 학술발표논문집 (KCC 2026)}, year = {2026} } ``` --- ## License - **Benchmark tasks, evaluation code, and results:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) - **Spatial datasets:** 공공누리 1유형 (출처 표시 조건, 각 데이터 출처 참조) - **QGIS MCP tool:** Not included. See [jjsantos01/qgis_mcp](https://github.com/jjsantos01/qgis_mcp)