INNOVATION AREAS

Sensors are a quickly evolving area of research.  
Within the ScaleAgData project we will work on the development and integration of an innovative sensing approach to detect sprayed pesticides 
Optimised use of pesticides and monitoring their use is among the core objectives of recent EU policies. ScaleAgData aims to eliminate the lack of monitoring instruments and suitable technologies for in-the-field automated detection of pesticides due to technical and cost issues. The use of compact, cost-effective low power and highly sensitive sensors is the main alternative to be explored. 
Another goal is to investigate whether specific conditions could be measured with hyperspectral sensors and reflectance-based technologies, considering the large amount of data, which poses a challenge when transferring the data. A hyperspectral imaging instrument will be customized to accommodate use-cases and data collection needs. The instrument's control and data interface will be designed to support the data collection and edge-computing objectives of the project. 

ScaleAgData will provide a platform for edge processing including AI functions deployment on the far edge. The focus is on ‘edge analytics’ to leverage the power of distributed edge computing to support real time stream processing at the data source. ScaleAgData will enable the dynamic composition, deployment and migration of data, data processing and data analytics as AI microservices. For model compression and reduction of the inference time, structured model pruning using PyTorch or Tensorflow will be used, also accelerating the model for each hardware device. 

ScaleAgData aims the development of a blueprint architecture for the collection of heterogeneous data from various devices, towards the creation of common datasets at regional, national and EU level.  
On top of this architecture, detailed designs will be developed identifying all the needed interactions between the architectural components for building each of the cases. 
ScaleAgData will contribute to the semantic technologies, data governance and interoperability for both streaming and repository-based (static) data. Semantic algorithms for data harmonization, verification, enrichment via semantic tagging, etc. will be developed with a clear focus on the needs of the agri-sector. These algorithms will use existing and newly created agri-related semantic models, and IoT-related ontologies, enabling efficient data management within IoT ecosystems, including application of AI algorithms. With the developed models, comprehensive practical assessment of usability of semantic technologies within edge-cloud continuum ecosystems will be completed. 

Due to a general lack of high-resolution training data, most image fusion activities have been carried out at low resolution and use original images as reference to evaluate their reconstruction accuracy. 
There is a lack of studies that could confirm that the proposed fusion techniques are able to maintain the initial high-resolution information during fusion.  
Highly detailed ground truth data acquired by in-situ sensors can present a viable alternative for evaluation of image fusion techniques in the absence of high-resolution imagery.  
ScaleAgData aims to develop generic DL fusion techniques that can adapt to diverse types of complementary image data sources covering a wide range of spatial, spectral, and temporal resolutions and will specifically focus on the data scarcity issue by involving in-situ sensor data within the process. 

ScaleAgData aims to develop methods for combining multiple models and sensor data sources into Prescriptive Digital Twins, enabling multi-objective decision-making on management actions using intelligent agents.  
Methods for automatically initializing and calibrating well-known simulation models will be developed based on a combination of existing data sources.  
Methods, software, and guidelines for the use of synthetic data for training deep neural networks capable of predicting e.g., crop fertilization requirements and water stress in scenarios where limited data is available, will be developed.  

Synthetic data will be generated using existing crop models across various soil types and historical weather conditions. The models trained with synthetic data and/or with transfer learning using real measured data, will be explored.  
Measurement data, ML (Machine Learning) models and PBMs will be integrated into digital twins in several case studies. 

ScaleAgData will gather data sharing privacy preserving user requirements in the co-design workshops to understand which techniques are more relevant to develop in this activity. 

One of the most promising methods is Federated Learning. It can be defined as a setting where several machines (clients) have data that cannot be shared, and a central entity (a server) coordinates the updates of the models that are trained individually in each client and aggregated in a central server.  
This setup also allows the global model to be partly retrained for a specific region with local data. 
Based on the user requirements analysis, which will be continuously updated, the project aims to develop a Privacy Preserving software library that would expose implementations of the different defined PET (Privacy Enhancing Technologies) techniques and could be tested in the different RI Labs. 

Many new and promising techniques in the fields of data assimilation and DL are emerging wich would enable a better integration between in-situ sensor data and existing spatially explicit data products. Many of these techniques have not been tested yet specifically in the context of agricultural monitoring, where many data products are typically spatially uncorrelated (e.g., crop type, Leaf Area Index, crop biomass and yield).  
Within the ScaleAgData project, we will provide the necessary technical frameworks to test the approaches, allowing the proper identification of the merits and limitations of each approach in the context of agricultural monitoring.