@soil-ai-lab/fedler-farms

🌱 FedLer-Farms

Federated Learning for Soil Property Prediction.

Associated Publication

This Flower application is aligned with:

Federated earth-observation models for collaborative farm-scale soil mapping
International Journal of Applied Earth Observation and Geoinformation, 146, 105067.

Gallios, G., Demattê, J.A.M., Tsakiridis, N., Cardoso, M.C., Kritharoula, A., Tziolas, N. (2026) (Link)

Overview

Accurate, privacy-preserving soil information is critical for:

Site-specific nutrient management
Carbon accounting
Sustainable agricultural practices

However, laboratory soil analyses remain expensive, resulting in sparse sampling grids at the farm level.

Our project demonstrates a federated learning (FL) framework that enables collaborative soil mapping across distributed farms without sharing raw soil data.

The approach:

Uses Sentinel-2 derived features (bare-soil composites)
Applies a 1D Convolutional Neural Network (CNN)
Trains collaboratively using Federated Averaging (FedAvg)
Keeps all local soil data on-premise

Model Architecture

Implemented in:

fedler_farms/model.py

Architecture:

1D Convolutional layer
ReLU
1D Convolutional layer
ReLU
Optional max pooling
Fully connected layer (64 units)
Multi-output regression head

Designed for spectral or tabular Sentinel-derived features.

Dataset

Demo dataset hosted on Hugging Face:

soil-ai-lab/dummy-soil-dataset

The dataset includes:

Feature columns (X1–X10)
Targets:
- Clay_gkg_filtered
- C_gkg_filtered

⚠️ Note: This is a simplified demonstration dataset.

The full experimental archive described in the paper is not publicly distributed.

Installation

python -m venv fedler-env
source fedler-env/bin/activate
pip install --upgrade pip
pip install -e .

Running (Simulation Engine)

Run fully local federated simulation:

flwr run . local-simulation

This will:

Spawn virtual clients
Partition the dataset
Train the federated CNN
Log metrics

Simulation Configuration

The local-simulation runtime is defined in Flower configuration file: ~/.flwr/config.toml.

Example:

[superlink.local-simulation]
options.num-supernodes = 3

This configuration runs the federated simulation locally with 3 virtual SuperNodes (clients).

Running (Deployment Engine)

This mode simulates real distributed farms.

Step 1 — Start SuperLink

flower-superlink --insecure

Step 2 — Start SuperNodes (Clients)

Example with 3 farms:

flower-supernode --insecure \
    --superlink 127.0.0.1:9092 \
    --clientappio-api-address 127.0.0.1:9104

flower-supernode --insecure \
    --superlink 127.0.0.1:9092 \
    --clientappio-api-address 127.0.0.1:9105

flower-supernode --insecure \
    --superlink 127.0.0.1:9092 \
    --clientappio-api-address 127.0.0.1:9106

Step 3 — Launch Federated Run

flwr run . local-deployment --stream

Where local-deployment is defined in config.toml:

[superlink.local-deployment]
address = "127.0.0.1:9093"
insecure = true

Metrics

Per round:

Centralized loss
Distributed loss
R² per target
RMSE per target
RPIQ per target

Outputs saved to: outputs/metrics_demo/.

Citation

If you use this application, please cite:

@article{Gallios2026FedEO,
  title   = {Federated earth-observation models for collaborative farm-scale soil mapping},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  volume  = {146},
  pages   = {105067},
  year    = {2026},
  doi     = {10.1016/j.jag.2025.105067},
  author  = {Gallios, G. and Demattê, J.A.M. and Tsakiridis, N. and Cardoso, M.C. and Kritharoula, A. and Tziolas, N.}
}

🔗 Link: https://doi.org/10.1016/j.jag.2025.105067

Developed by

Soil Science Artificial Intelligence Laboratory

University of Florida