A team of researchers has developed a knowledge-guided machine learning algorithm that uses satellite measurements to provide much more precise estimates of soil moisture across the globe. The researchers hope the data can eventually help farmers make better decisions about when to sow crops or irrigate fields, explains one of the scientists behind the project.
Soil moisture has become a kind of invisible water meter for our planet.
It reveals whether fields have enough water to support crop growth – or whether they are becoming too dry or too wet for healthy development. If the soil becomes too dry, crops can experience water stress long before the damage becomes visible in the plants.
The challenge is that soil moisture is difficult to measure at large scale – precisely where decisions need to be made. This is where the new algorithm comes in.
“Our aim is to help farmers and other decision-makers make better decisions based on more accurate data. This represents a significant advance compared with the options that were previously available,” says a researcher behind the project, Sheng Wang, Associate Professor from the Centre for Landscape Research in Sustainable Agricultural Futures at Department of Agroecology, Aarhus University in Denmark.
The algorithm and the research behind it have been published in Nature Scientific Data.
When the soil dries out, the whole system feels it
Interest in the water content of soil is nothing new. Soil moisture is not only crucial for crop yields but also for the global water, carbon and energy cycles.
As climate change reshapes these cycles across the globe, new kinds of problems are beginning to emerge.
In 2018 and 2020, severe drought struck large parts of Europe, costing farmers dearly as crops withered and harvests failed. Soil moisture is closely linked to drought impacts because it reflects how much water is available for plants.
“Traditionally, soil moisture is measured locally using sensors placed in the ground – a method that is both expensive and limited in coverage,” explains Sheng Wang.
Both the United States space agency NASA and its European counterpart ESA have enabled soil moisture to be measured indirectly from space.
This is done using passive microwave signals, which measure radiation emitted from the Earth’s surface. The signals are sensitive to water close to the surface and can penetrate clouds, making them useful for large-scale monitoring.
From satellite signals to real measurements of soil moisture
The signals reflected from the Earth’s surface contain information about how much water the soil holds – but they must be interpreted through models to be turned into concrete measurements.
However, the method is not entirely accurate. Vegetation, soil properties, surface temperature and terrain can interfere with the signal, making it harder to estimate how much water is available in the soil – especially in vegetated areas.
It is precisely this challenge that the new algorithm seeks to address.
By combining artificial intelligence with satellite data and around half a million ground-based measurements collected over 10 years across the globe, the team developed a model that can interpret the signals much more accurately – and has tested it against independent measurements taken directly in the soil.
The knowledge-guided machine learning algorithm combines physical understanding of microwave radiative transfer with ground-based measurements to improve soil moisture estimates across many different climates and landscapes. In previous studies, the algorithm has also been used to study crop carbon cycling, aboveground biomass and nitrogen uptake.
“We combine our physical understanding of the satellite signals with artificial intelligence so that we can both explain and improve the estimates,” says Sheng Wang.
New model sharply improves on existing satellite estimates
The new algorithm outperforms existing soil moisture data products from ESA and NASA – even when evaluated against the same ground-based measurements.
Performance was evaluated using several metrics, including the correlation coefficient (R), on a scale from 0 to 1, where 1 indicates very close agreement with measured soil moisture.
“In the evaluation, the new algorithm reaches around 0.9, showing stronger agreement with ground measurements than several widely used satellite-based soil moisture products. This marks a clear improvement, including in areas where satellite-based estimates have traditionally had high uncertainty,” says Sijia Feng, the first author of this study and a PhD student at Aarhus University.
Feng co-led the study alongside Aoyang Li, an undergraduate student at Agroecosystem Sustainability Center and National Center for Supercomputing Applications, University of Illinois Urbana-Champaign.
“To make the data easier to access, we have also developed high-resolution soil moisture maps and an interactive platform. The data can be accessed on a mobile phone or laptop. Farmers will be able to monitor soil moisture and adjust their farming practices accordingly,” he concludes.
