The instrument at the heart of the study is the Advanced SCATterometer, or ASCAT, a C-band radar flying aboard the Metop series of polar-orbiting satellites operated by the European Organisation for the Exploitation of Meteorological Satellites. ASCAT works by bouncing microwave pulses off the Earth’s surface and measuring the strength of the returned signal, known as backscatter. Because water dramatically changes the dielectric properties of soil, wetter ground reflects radar differently than dry ground, and this relationship can be exploited to estimate how saturated the top roughly five centimeters of soil are. Three Metop satellites — launched in 2006, 2012, and 2018 — orbit at about 817 kilometers altitude, giving near-global coverage every twelve hours and typically observing any given location two to four times per day depending on the orbit geometry.

That sounds frequent, but it poses a real problem for capturing sub-daily dynamics. Most existing soil moisture products, including the operational ASCAT H120 climate data record used as a benchmark in this study, are evaluated and used at daily resolution, because the retrieval uncertainties are large relative to the small changes in moisture that occur between overpasses. The signal-to-noise ratio is simply too low to trust hourly estimates. Dinh and colleagues asked whether a machine learning approach could push past this barrier, and their answer was a localized convolutional neural network, or CNN-l, that departs from the way most satellite retrievals have traditionally been built.

The key innovation lies in how the network treats space. Conventional neural network retrievals operate pixel by pixel, ignoring the fact that satellite observations carry rich spatial context — neighboring landscapes, shared climate regimes, and coherent terrain features all carry information about soil moisture. Convolutional neural networks, originally developed for image recognition, excel at extracting such spatial patterns through learnable filters that slide across the input. The team went one step further by using locally connected layers, in which each spatial location gets its own set of filter weights rather than sharing weights across the whole domain. This allows the model to adapt to regional conditions — the vegetation density of the southeastern United States, the complex topography of the mountainous West — rather than forcing a single global set of rules onto radically different landscapes.

The model was trained over the contiguous United States, a region chosen for its pronounced climatic diversity and its dense network of ground-based monitoring stations. The inputs were physically motivated: ASCAT’s normalized backscatter at a 40-degree reference incidence angle, plus two auxiliary variables from the ERA5 reanalysis produced by the European Centre for Medium-Range Weather Forecasts — soil temperature in the top seven centimeters, which affects the thermal environment of the soil, and leaf area index, which captures how much vegetation attenuates the radar signal. The network learned to predict volumetric surface soil moisture against the ERA5 target using data from 2016 to 2018, with 2019 completely held out for independent testing. Separate models were trained for ascending and descending satellite passes, and overfitting was controlled through early stopping that halted training when validation performance plateaued.

The results against the ERA5 reference were striking. The CNN-based retrievals achieved total correlation coefficients exceeding 0.9 at the sub-daily scale — 0.92 for ascending passes and 0.91 for descending passes — compared with roughly 0.58 and 0.59 for the operational H120 product. But because the model was trained against ERA5, agreement with that reanalysis is only part of the story. The more demanding test came from 568 in situ monitoring sites across the United States, drawn from the Soil Climate Analysis Network, the SNOTEL network, and the United States Climate Reference Network, all part of the International Soil Moisture Network. These ground measurements were entirely withheld from training, providing a genuinely independent yardstick.

Against those ground stations, the CNN retrievals achieved a median temporal correlation of 0.65, comfortably ahead of the 0.59 achieved by the operational H120 product, and with error levels approaching those of ERA5 itself. The improvements were not uniform: the deep learning approach won at 352 of the 568 sites, about 62 percent, with particularly strong gains in the western and southeastern regions and over forested, mountainous terrain where spatial context matters most. The operational product retained a slight edge over croplands in the central Great Plains, and it showed lower median bias — though the authors note this reflects local calibration rather than superior skill in tracking moisture dynamics. Notably, despite being trained on ERA5, the CNN outperformed its own training target over crops and grasslands when judged against ground truth, suggesting the satellite observations were contributing genuine physical information beyond the reanalysis background.

The most tantalizing findings concern intraday variability — the question of whether soil moisture changes within a day can be detected at all from a satellite that visits only a few times daily. In a detailed case study from 20 November 2019, the team examined a heavy precipitation event in the American Southwest, cross-checking the retrieved moisture response against the independent Multi-Radar Multi-Sensor gauge-corrected precipitation product. As rain began around 04:00 UTC and continued through the day, the CNN retrievals from three ASCAT overpasses tracked a rise in soil moisture consistent with both the reanalysis and the observed rainfall, capturing a daily amplitude of about 0.2 cubic meters per cubic meter. Crucially, the alignment with independent precipitation data confirmed the response was physically real, not merely an echo of the ERA5 forcing used in training.

Scaling up to the full year 2019, the team compared diurnal moisture amplitudes across more than 4.4 million pixel-days. Without filtering, the correlation with the reanalysis was a modest 0.21 — most of the apparent variation was retrieval noise. But as precipitation thresholds were applied, the signal emerged from the noise. On days with rainfall exceeding 10 millimeters and no rain the previous day, correlation rose to 0.47, and the model captured daily maximum and minimum moisture values with correlations of 0.82 and 0.88 respectively. The pattern is physically intuitive: intraday signals become detectable precisely when they are largest — during intense rain falling on relatively dry soils — and the amplitude of the moisture swing exceeds the retrieval uncertainty. The model does tend to dampen extremes, a well-known tendency of statistical regression approaches, underestimating peak wetting and slightly overestimating the driest values.

The authors are candid about the limitations. With only a handful of overpasses per day, the exact timing of daily moisture maxima and minima is constrained to the satellite’s visit windows, and the full diurnal cycle remains out of reach for a single instrument. Their prescription is multi-sensor integration: fusing ASCAT with missions such as SMOS and SMAP, which sense at different times and frequencies, could dramatically improve both temporal resolution and accuracy, though it introduces challenges of inter-sensor calibration and heterogeneous uncertainties. They also suggest that adding soil texture, precipitation information, and dynamic model components could sharpen sensitivity to rapid changes, and that emerging foundation models in remote sensing may help weave heterogeneous data sources into consistent retrievals. The implications stretch well beyond soil moisture science. Sub-daily moisture monitoring could sharpen flood forecasting when soils saturate rapidly, improve agricultural irrigation decisions, and feed better land-atmosphere coupling information into weather prediction. And the localized CNN framework itself is portable: the authors point to applications for land surface temperature, surface emissivity, vegetation cover, and surface water extent — any geophysical variable with strong spatial structure that satellites observe imperfectly. What was once a daily snapshot of the water held in the ground may soon become something closer to a movie.

Subject of Research: Deep learning retrieval of sub-daily surface soil moisture from ASCAT satellite observations

Article Title: Evaluation of ASCAT soil moisture retrievals and their potential to detect intraday variability

Article References: Dinh, L. A., Aires, F., & Pellet, V. (2026). Evaluation of ASCAT soil moisture retrievals and their potential to detect intraday variability. Earth Observation, 1(1), 105-120. https://doi.org/10.5194/eo-1-105-2026

Image Credits: AI Generated

DOI: 10.5194/eo-1-105-2026

Keywords: soil moisture, ASCAT, deep learning, convolutional neural networks, remote sensing, satellite retrieval, ERA5 reanalysis, intraday variability, hydrology, Metop satellites, flood forecasting, International Soil Moisture Network