DSSAT-MCP: AI Agent Interface for Crop Model Calibration

A proof-of-concept MCP (Model Context Protocol) server that wraps the DSSAT CSM v4.8 crop simulation model, enabling natural language interaction through LLM agents (e.g., Claude).

Paper: "MCP-Based AI Agent Interface for Crop Model Calibration: A Proof of Concept with DSSAT" (under review)

Overview

Crop model calibration traditionally requires deep technical expertise: writing experiment files, configuring parameters, parsing outputs. This project demonstrates that wrapping DSSAT in an MCP server allows an LLM agent to perform calibration tasks through natural language alone.

Available Tools (9 MCP tools)

Supported Crops

Related MCP server: ladybug-tools-mcp

Prerequisites

1. DSSAT v4.8

Download and install from dssat.net (free registration required). Default install path: C:\DSSAT48

2. Python 3.10+

pip install -r requirements.txt

3. MCP-compatible client

• Claude Desktop — recommended

• Any MCP-compatible LLM client

Installation

Step 1: Clone this repository

git clone https://github.com/YOUR_USERNAME/dssat-mcp.git
cd dssat-mcp

Step 2: Install Python dependencies

pip install -r requirements.txt

Step 3: Copy data files into DSSAT

# Weather files
copy data\SUWO2501.WTH  C:\DSSAT48\Weather\
copy data\SUWO2601.WTH  C:\DSSAT48\Weather\

# Soil profiles
copy data\KR.SOL        C:\DSSAT48\Soil\

# Korean cultivar parameters
copy genotype\WHCER048.CUL  C:\DSSAT48\Genotype\WHCER048.CUL
copy genotype\BACER048.CUL  C:\DSSAT48\Genotype\BACER048.CUL
copy genotype\MZCER048.CUL  C:\DSSAT48\Genotype\MZCER048.CUL
copy genotype\SBGRO048.CUL  C:\DSSAT48\Genotype\SBGRO048.CUL

Note: The CUL files in genotype/ contain Korean cultivar entries added to the original DSSAT files. Back up your originals before copying.

Step 4: Configure environment variables

copy .env.example .env
# Edit .env with your paths

Step 5: Register with Claude Desktop

Edit %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "dssat-mcp": {
      "command": "python",
      "args": ["C:/path/to/dssat_mcp_server.py"],
      "env": {
        "DSSAT_HOME":    "C:/DSSAT48",
        "DSSAT_BIN":     "C:/DSSAT48/DSCSM048.EXE",
        "DSSAT_WORK":    "C:/dssat_jobs",
        "DSSAT_WEATHER": "C:/DSSAT48/Weather",
        "DSSAT_SOIL":    "C:/DSSAT48/Soil"
      }
    }
  }
}

Quick Start

Once Claude Desktop is running with the MCP server registered, you can interact naturally:

"Run a maize simulation for Suwon, sowing May 1 2025, 120 kg N/ha"

"Estimate cultivar parameters for Korean wheat sown Oct 25 2025,
 heading date Apr 20 2026, maturity Jun 7 2026, yield 5200 kg/ha,
 thousand grain weight 38g"

"Compare nitrogen rates 0, 60, 120, 180, 240 kg/ha for maize
 at Suwon using sensitivity analysis"

"Evaluate simulation accuracy:
 observed yield 5000, simulated 4919;
 observed yield 6200, simulated 5850"

Data Files

Weather (data/)

Soil (data/)

Genotype (genotype/)

Korean cultivar parameters added to standard DSSAT .CUL files:

• KR0003 — Maize Dacheongok (옥수수 다청옥)

• KR0001, KR0002 — Wheat Tapgol / Seodunchal (밀 탑골/서둔찰)

• KR0001, KR0002 — Barley Tapgol / Seodunchal (보리 탑골/서둔찰)

• KR2828 — Soybean KRUG2828 (콩)

Key Features

Calibration (estimate_cultivar_params)

Estimates DSSAT cultivar parameters directly from field observations — no optimization loop required:

• P5: GDD from heading to maturity (all crops)

• P1V: Vernalization days (wheat, barley)

• G2/G3: Kernel weight from thousand-grain weight

• Verification simulation run automatically after estimation

Model Evaluation (evaluate_simulation)

Standard statistical metrics for model performance assessment:

• RMSE, MAE, MBE (bias)

• Willmott d-index

• Nash-Sutcliffe Efficiency (NSE)

• Pearson R²

Climate Scenarios (run_batch)

# Example: RCP scenario comparison
run_batch(crop="maize", scenarios=[
    {"label": "baseline",  "sowing_date": "2025-05-01"},
    {"label": "+2°C",      "delta_temp": 2},
    {"label": "RCP4.5",    "delta_temp": 2, "co2_ppm": 550},
    {"label": "RCP8.5",    "delta_temp": 4, "co2_ppm": 700},
])

Limitations

• Single AI agent (Claude) — extensible to other LLM clients via MCP protocol

• Single crop model (DSSAT) — architecture supports adding APSIM, STICS, etc.

• Definition-based calibration: accurate for phenology parameters (P5, P1V), less so for yield parameters without anthesis biomass data

• Windows native; Linux/macOS require Wine

System Architecture

User (natural language)
        │
        ▼
  LLM Agent (Claude)
        │  MCP Protocol (JSON-RPC over stdio)
        ▼
  DSSAT-MCP Server (Python / FastMCP)
        │
        ├── FileX writer (experiment file)
        ├── Weather handler (perturbation, KMA download)
        ├── Soil handler (profile lookup)
        ├── Cultivar estimator (parameter estimation)
        └── Output parser (Summary.OUT, PlantGro.OUT, ...)
              │
              ▼
         DSSAT CSM v4.8 (DSCSM048.EXE)

Citation

If you use this code, please cite:

@article{yourname2025dssat,
  title   = {MCP-Based AI Agent Interface for Crop Model Calibration:
             A Proof of Concept with DSSAT},
  author  = {Your Name et al.},
  journal = {Computers and Electronics in Agriculture},
  year    = {2025},
  note    = {under review}
}

License

MIT License — see LICENSE for details.

DSSAT itself is subject to its own license agreement (dssat.net).