Predicting Water Flow Intensity with Machine Learning

Moreover, the historical drainage of wetlands and straightening of streams have disturbed the natural water cycle and exacerbated the effects of extreme weather events.

To cope with these challenges, we need to understand how water flows in the environment and how it is influenced by various factors, such as rainfall, temperature, land cover, soil, and elevation. This can help us make informed decisions about nature-based climate adaptation techniques, such as restoring wetlands, greening urban areas, and protecting soil. However, traditional hydrological models that rely on expert knowledge and physical properties are limited in their ability to generalize and capture the complexity and variability of water flow dynamics.

In this project, we propose a novel machine learning (ML) approach for predicting water flow intensity that leverages both temporal and spatial data. We use a fully convolutional neural network (FCN) that receives spatio-temporal inputs and predicts the water flow intensity at every coordinate for the next day. Our approach has two main advantages:

• We include spatial data as inputs to the model, such as satellite imagery and several derived geographic information system (GIS) layers. This allows the model to learn the relationships between temporal and spatial aspects of the local environment, such as how land cover, soil depth and moisture, and elevation affect water flow.
• We tackle the task of dense water flow intensity prediction, meaning that we predict the water flow intensity for every pixel in the image, rather than only for a few selected locations. This way, we can obtain a more detailed and comprehensive picture of the water flow dynamics in the area.

Project Results

The approach has been evaluated on a dataset of water flow intensity measurements from catchment areas in Sweden, covering different climatic and environmental conditions. The evaluation included ablation studies and analysis of the improtance of different factors.

The approach shows strengths in performance, is highly flexible in the information sources that can be provided as inputs to the model, and gives a dense prediction with flow estimations in every point of the input map. We also show some qualitative examples of the predictions made by our model, which illustrate how the model captures the spatial patterns and variations of water flow intensity in different catchment areas.

Project Implications

Our project demonstrates the potential of using ML for water flow intensity prediction, which can provide valuable information for climate adaptation and water management. By using both temporal and spatial data, our model can learn the complex and dynamic interactions between water flow and the environment, and produce high-resolution predictions that can reveal the spatial distribution and heterogeneity of water flow intensity. This can help us better understand the flood risks and the effects of flood and drought mitigation, as well as the general hydrological implications of changes in land use.

Our project also opens up new directions for future research and development, such as:

• Extending our approach to other time horizons, such as hourly or weekly predictions, or longer-term projections.
• Incorporating other types of data, such as groundwater or snow cover, to improve the accuracy and robustness of the model.
• Exploring the explainability and interpretability of the model, to understand how the model makes predictions and what features it relies on.
• Developing user-friendly and interactive tools and interfaces, to facilitate the communication and visualization of the model predictions and their uncertainties.