TerrAI

Source code for the paper "Towards Data-driven Nitrogen Estimation in Wheat Fields using Multispectral Images"

Table of Contents

• Overview
• Prerequisites
• Installation
  • Using Docker Compose
  • Manual Python Environment
• Data Preparation
  • Parsing raw data
  • Weather aggregation
  • Patch creation & preprocessing
• Model Training, Inference & Visualization
• Contributors
• Acknowledgement

Overview

TerrAI is an efficient CNN-based solution for Targeted Spraying and Fertilization (TSF) over wheat fields. It ingests raw field data (Excel sheets, shapefiles, GeoTIFFs), enriches it with weather information, creates raster patches, trains a UNet model, and produces visualizations of forecasted prescriptions.

Prerequisites

• Docker (>=28) – for creating the experimental testbed.
• Docker Compose (>=v2.34.0) – for automating container creation.
• GPU (optional) – if you want to accelerate training with CUDA, modify the Dockerfile accordingly.
• URL and API token for the weather service (required for weather-data.py).

Installation

1. Clone the repository.
2. Build and start the service.

docker-compose up --build -d

The docker-compose.yaml defines the service (TerrAI testbed) that builds the image from the provided Dockerfile and mounts the src directory as a volume, giving you live access to the code inside the container.

Data Preparation

Parse field data

python src/data-parsing.py

Weather forecasts

python src/weather-data.py --url <URL> --token <TOKEN>

Patch creation & pre-processing

python src/data-preprocessing.py \
    --spectral_bands BNDVI \
    --weather_bands "min T; T0" \
    --weather_bands "min T; T8" \
    --weather_bands "min T; T16" \
    --weather_bands "mean rel. hum.; T0" \
    --weather_bands "mean rel. hum.; T8" \
    --weather_bands "mean rel. hum.; T16" \
    --weather_bands "mean pressure; T0" \
    --weather_bands "mean pressure; T8" \
    --weather_bands "mean pressure; T16" \
    --weather_bands "max gust; T0" \
    --weather_bands "wind direction; T0" \
    --weather_bands "wind speed; T16" \
    --weather_bands "global energy; T0" \
    --kernel 8 \
    --stride 1 \
    --lands Wheat \
    --phases 2

Model Training, Inference & Visualization

The high‑level training script is src/training-terrai-v10.py. In our experimental study, we used the following command for training the baseline TerrAI model

python training-unet-v10-raster-patch.py \
    --spectral_bands BNDVI \
    --weather_bands "min T; T0" \
    --weather_bands "min T; T8" \
    --weather_bands "min T; T16" \
    --weather_bands "mean rel. hum.; T0" \
    --weather_bands "mean rel. hum.; T8" \
    --weather_bands "mean rel. hum.; T16" \
    --weather_bands "mean pressure; T0" \
    --weather_bands "mean pressure; T8" \
    --weather_bands "mean pressure; T16" \
    --weather_bands "max gust; T0" \
    --weather_bands "wind direction; T0" \
    --weather_bands "wind speed; T16" \
    --weather_bands "global energy; T0" \
    --kernel 8 \
    --stride 1 \
    --land Wheat \
    --phase 2 \
    --use_augmentation \
    --gpuid 0 \
    --hidden_channels 24,36,48 \
    --dropout 0.25

Adjust the --hidden_channels parameter to 4/8/16 and 72/84/96 for training the small and large TerrAI variants, respectively. After training, use the following methods from src/helper.py

• model_inference - folds model predictions back to full‑size rasters via st_image_toolkit.unary_union.
• visualize_model_results - creates side‑by‑side plots of NDVI, ground‑truth prescription, and model prediction with colorbars.

Contributors

• Andreas Tritsarolis; Department of Informatics, University of Piraeus
• Tomaž Bokan; Innovation Technology Cluster
• Matej Brumen; Department of Electrical Enginnering & Computer Science, University of Maribor
• Domen Mongus; Department of Electrical Enginnering & Computer Science, University of Maribor
• Yannis Theodoridis; Department of Informatics, University of Piraeus

Acknowledgement

This work was supported in part by the Horizon Europe Research and Innovation Programme of the European Union under grant agreement No. 101070416 (Green.Dat.AI; https://greendatai.eu). The authors would also like to acknowledge INESC-TEC for providing access to their Energy benchmarking tool.