NiMo4.0

Challenge

The NiMo 4.0 project has developed a system that combines AI processes with environmental informatics methods, specifically within the field of water. To enable the effective use of various data sources and AI methods in operational settings, new developments in intelligent sensor data transmission and processing were made practical and implemented in various end-user platform demonstrators based on the new open standard Sensor-ThingsAPI from the Open Geospatial Consortium (OGC).

The project's main goal was to enhance the spatial and temporal prediction of nitrates in groundwater and develop intelligent decision support systems based on these predictions. These systems could contribute to optimising groundwater protection programmes and efficiently reducing nitrates through scenario calculations, making the process more sustainable. These approaches and methods were developed, demonstrated and validated using real data from two important water management regions with sufficient hydrogeological variability to enable conclusions to be drawn about the general validity and transferability of the developed solutions. Additionally, spatial prediction in conjunction with modern geostatistical and operational research methods enables recommendations for optimising measurement networks.

Motivation

The distribution of nitrate in groundwater is a fascinating and complex phenomenon influenced by various factors. It is the result of the intricate interplay between land use, meteorological conditions, the chemical and physical properties of soil layers, and transport and reaction processes within the groundwater itself. This diversity of factors gives rise to highly complex and dynamic nitrate distribution patterns in groundwater that can vary greatly from region to region and vertically.

Understanding and modelling this complex system is a significant challenge. Previous analytical and numerical models are unable to accurately map the spatial and temporal variability of nitrate content in groundwater. This is where AI applications, particularly machine learning and deep learning methods, can help. These innovative technologies have the potential to extract and interpret complex patterns and relationships from large amounts of data.

Using AI allows us to gain new insights by delving deeper into the structure and dynamics of nitrate distribution in groundwater. These data-driven models allow us to identify hidden correlations and make accurate predictions about nitrate spread in groundwater. AI helps us to better understand the complexity of this hydrogeochemical system, enabling us to develop more effective measures to reduce nitrate pollution.

Solution

The system architecture was to be specified by the Fraunhofer IOSB based on open standards. This was to be done in collaboration with the project coordinator, disy, and the implementation was to be carried out by the Fraunhofer IOSB for intelligent data management. To this end, the Fraunhofer IOSB further developed the open-source FROST® implementation of the OGC SensorThings API.

In NiMo 4.0, the first FROST server® was set up using nitrate data from the pilot regions of Baden-Württemberg and Lower Saxony. This data was then made available directly to the other NiMo 4.0 partners (AGW, TZW and the UWR department of IOSB-AST) for the development, training and execution of the various ML algorithms to be researched in NiMo 4.0 via the STA interface. As the ML methods in NiMo 4.0 were implemented in Python, a new client library for the Python client was developed for the FROST server®, which is now also available as open source software. During the project, other state environmental agencies expressed interest in the developments and provided their own data. This enabled the creation of a data server containing a cross-state treasure trove of data comprising over 6 million nitrate measurements and nitrate-related chemical parameters, such as groundwater level, conductivity, temperature, pH value, oxygen and turbidity. As part of NiMo 4.0, the real-time connection of data loggers for nitrate measurements and the continuous integration of the data into the FROST Server® were also successfully tested. 

In addition to its data access functions, the SensorThingsAPI offers a 'tasking' extension that allows tasks to be carried out via the STA's standard interface. This interface was first tested in NiMo 4.0 for executing ML methods. To this end, the PERMA implementation was further developed to encapsulate ML models in a Docker container and execute them via the STA.

Fraunhofer IOSB has successfully implemented two further tasks. Firstly, the usability of convolutional neural networks for regionalising nitrate was investigated. The machine learning models were programmed in Python and tested in a case study in Baden-Württemberg. Secondly, new methods based on deep learning were developed and tested to identify anomalies and events in multivariate spatiotemporal groundwater data. A case study was conducted for Lower Saxony and Thuringia.