Welcome to the 4th edition of our ScaleAgData Newsletter! We're excited to continue empowering modern agriculture with secure, collaborative AI technologies.
In this Newsletter we explore how Federated Learning is transforming agricultural data sharing, balancing privacy with the power of large, decentralized datasets. We also highlight our latest experiment in the Soil Health Research and Innovation Lab, where EV ILVO and AUTh are developing a decentralized topsoil Soil Organic Carbon model using federated learning, advancing soil research and supporting the EU Soil Strategy for 2030. Thank you for being part of this journey. Together we can drive innovation while respecting data sovereignty for a more sustainable agricultural future. Enjoy!
Stay tuned via our website & social media channels & don't hesitate to forward this newsletter or invite colleagues to register!
In this 4rd ScaleAgData Newsletter:
6th innovation area: Privacy preserving technologies for modern agriculture
End-user quotes
Recent News items
Upcoming events
Intro
In modern agriculture, data is a valuable asset, often kept private to both protect knowledge and ensure privacy, while creating barriers to collaboration and innovation. Traditional machine learning relies on centralized data, making it difficult to ensure security and meet regulations. Yet modern AI systems need diverse, large-scale datasets to develop robust, scalable solutions for regional and global challenges. The ScaleAgData project is tackling this with federated learning, enabling collaborative model training across decentralized datasets without compromising privacy.
Our aim: empower agriculture with secure, scalable AI solutions that respect data sovereignty and unlock the power of shared knowledge.
Innovation area 6: Privacy preserving technologies for modern agriculture
Federated Learning is a concept in artificial intelligence that enables multiple devices or servers to collaborate on training a machine learning model without sharing raw data. Instead, each device trains the model locally using its own data and sends only the model parameters or updates to a central server. The server aggregates these updates to create a more robust global model.
In Earth Observation, Federated Learning (FL) enables collaborative training of machine learning models across distributed datasets without centralizing sensitive or large-scale remote sensing data. Satellite operators, research institutions, and environmental agencies can locally process data to train models, sharing only model updates with a central server. This approach is ideal for applications like monitoring crop growth, deforestation, tracking climate change, and assessing natural disasters, as it enhances data privacy, reduces bandwidth usage, and facilitates global cooperation. Federated Learning ensures that diverse datasets from different regions and organizations contribute to a robust global model while preserving the confidentiality of proprietary or sensitive geographic information.
In the context of the ScaleAgData project, a Federated Learning toolbox extending the capabilities of the flower.ai framework will be developed to support the decentralized collaboration among diverse stakeholders in the development of advanced AI models for agricultural and environmental applications in the different RILs. This approach allows, for instance, the creation of robust AI models for precision farming, crop yield prediction, and environmental impact assessments, while ensuring data privacy and security.
Flower.ai is an open-source framework designed to simplify and scale federated learning systems. It provides tools and infrastructure for building, deploying, and managing Federated Learning experiments across diverse environments, including edge devices, distributed servers, and cloud platforms. Flower.ai supports multiple machine learning frameworks, such as TensorFlow, PyTorch and JAX, and enables developers to implement custom FL strategies tailored to specific use cases.
Currently the development of the ScaleAgData Federated Learning toolbox is being driven by an experiment led by EV ILVO and AUTh in the Soil Health RIL which is focused on building a decentralised topsoil Soil Organic Carbon (SOC) model at European level based on models trained at national scale built with privately owned data. In the second part of the project, this component will include other ScaleAgData RILs, further evolve based on their requirements and will be shared with the general public via the ScaleAgData public Github repository.
"Federated Learning empowers Soil Living Labs to collaborate securely, sharing insights without compromising data privacy. This fosters innovation, accelerates solutions to soil degradation, and supports Europe's mission to restore soil health through collective action."
Evdokimov Konstantinidis Vice Chair of the European Network of Living Labs (EnoLL)
Soils are the basis for all terrestrial life and serve as a significant carbon sink. Today, an estimated 70% of EU soils are degraded, making their protection and restoration a key mission of the EU Soil Strategy for 2030, which establishes a framework with concrete measures to address this crisis, with the upcoming soil monitoring law representing a significant step toward achieving these objectives.
In soil science, however, while many organizations conduct soil analyses for farmers, land owners etc., the resulting data is often kept private because it is tied to specific farm geolocations-highly sensitive information. This presents a major challenge, as EO and other innovative technologies used for soil monitoring, depend on precise geolocations to correlate various geospatial input layers with the soil properties. Unfortunately, the reluctance to share data creates a significant barrier to advancing soil research and conservation efforts.
Federated Learning offers a transformative solution by enabling the training of neural networks and other machine learning models across decentralized datasets while preserving privacy. In our Research and Innovation Lab on Soil Health, we are using Federated Learning to develop robust, transferable models for soil property prediction without requiring direct data sharing. This approach could revolutionize the soil science community, fostering an unprecedented era of knowledge sharing and collaboration. By demonstrating the feasibility of secure, collaborative AI development, we aim to bridge the gap between privacy concerns and the need for comprehensive data analysis. This approach can be used in various scenarios, including developing models from laboratory infrared spectra or data from remote sensing (air- or space-borne).
Our initial efforts focus on soil data in croplands from Belgium's Flanders (EV ILVO) and Greece's Region of Central Macedonia (AUTh), building on a base model trained on the LUCAS topsoil database from the European Soil Data Centre (ESDAC) using Copernicus Sentinel-2 data. This foundation allows us to create scalable models with the potential for significant impact at a regional and national scale.
Looking forward, we plan to integrate hyperspectral data from upcoming missions like KUVA Hyperfield-1 and EnMAP, to further enhance model performance. By combining Federated Learning with advanced EO technologies, we aim to provide actionable insights that support soil monitoring and sustainable land management across Europe.
"Federated Learning enables collaborative training across decentralized soil spectroscopy datasets, overcoming privacy challenges. It fosters robust, generalizable models for accurate soil predictions, leveraging diverse regional data without sharing sensitive information."
Professor José A.m. Demattê University of Sao Paulo, Luiz de Queiroz, Brazil
The two study areas defined in the Grassland Research and Innovation Lab are very different in terms of geographical, meteorological, and orographic characteristics and vegetation management. Therefore, they represent two very distinct grassland scenarios. However, for both scenarios, a series of field measurements campaigns have been conducted over the past two years to measure biomass.
One of the objectives of the laboratory is to establish a relationship based on a statistical model between Leaf Area Index (LAI) measurements (obtained from various EO sensors) and in situ biomass measurements. The Flower AI Federated Learning model developed within the ScaleAgData project will be utilized to achieve this. This approach allows the data collected in the two sites to be processed in parallel while preserving data anonymity. The methodology will enable testing a learning model that allows any institution willing to contribute to the model's training to use its own in situ biomass data without sharing it. This model will be tested using LAI data derived from Sentinel-1 and Sentinel-2 satellite sensors. The tests will evaluate both the simple LAI point data corresponding to the biomass measurements and the time series of LAI at the measurement points. An analysis of the performance of the developed models will then be presented, laying the foundation for a monitoring service on a European scale.
Recent News
ScaleAgData at GEOGLAM Open Science Day
ScaleAgData made a significant contribution at the recent GEOGLAM Open Science Day, illustrating how access to in-situ sensor data can revolutionize EO monitoring capabilities in agriculture
Paper: Water stress on Mediterranean oak savanna grasslands
We're happy to share with you this paper on the impact of water stress on Mediterranean oak savanna grasslands.
Would you like to learn about upcoming events focused on the latest innovative data technologies for managing agricultural production and monitoring agricultural environments?
VITO is the project coordinator of the ScaleAgData consortium and will also contribute significantly to the 'product and service development', the RI Lab on Yield monitoring as well as the entire communication segment of the project. As a technology provider, VITO will bring data products and innovative solutions to merge sensor data into the data products. As communication lead, VITO will be managing the ScaleAgData website, social media channels as well as these digital newsletters. Feel free to contact us via scaleagdata@vito.be.
------------------------------------- If you don't want to receive our ScaleAgData newsletter, you can unsubscribe: see the 'unsubscribe' link below.
SHARING IS CARING
Feel free to share this newsletter with colleagues or on social media!
.jpg?upscale=true&width=1120&upscale=true&name=Grasslands_chang-qing-8oPubUm97Cc-unsplash%20(1).jpg)
