An adaptive and unsupervised approach for irrigation detection at field scale from high-resolution soil moisture maps

Monitoring irrigated areas and water requirements remains a key challenge in Earth Observation (EO), especially in regions experiencing growing water stress and agricultural intensification driven by rising demand and climate change [1-2]. An emerging methodology for detecting irrigated fields at large scale uses EO-derived high-resolution surface soil moisture maps (SSM). This approach can effectively segment irrigated and non-irrigated areas early in the season. Particularly, SSM derived from Synthetic Aperture Radar (SAR) data offer the resolution required to resolve irrigated fields and detect irrigation events, even before crop canopy development [3].

This study investigates the use of high-resolution (~100 m) SSM maps to detect irrigated fields in the Apulian Tavoliere agricultural district (Southern Italy), where winter cereals and tomato are the main cultivated crops. The SSM maps are derived from Sentinel-1, SAOCOM, and Sentinel-2 time series using the SMOSAR software developed at CNR-IREA [4]. The analysed data set covers the growing season 2024 and 2025. The irrigation detection is based on the application of the Constant False Alarm Rate (CFAR) algorithm. This methodology uses a sliding-window approach to classify the central pixel by comparing its value to a threshold derived from the probability distribution function of SSM values within the window, ensuring a fixed FAR. The result is the identification of fields showing higher SSM than their surrounding area. The probability distribution function adopted is the Gaussian Mixture, and the sliding window is a 3kmx3km square. Finally, the classified pixels are aggregated at the field scale using the parcel boundary information to evaluate the classification performance metrics.

Results indicate that the main factors affecting classification accuracy are satellite revisit time, vegetation stage, and radar frequency. Specifically, satellite revisit affects accuracy as SSM contrast decreases due to evapotranspiration, making detection challenging beyond three days after the irrigation. Furthermore, dense vegetation limits C-band SAR signal penetration into the soil, thereby ensuring detection is most effective during early crop growth. Analysis of the 2024 season shows that, at the start of growth, accuracy reaches 80%. While, at C-band, as vegetation matures, the canopy may dominate the backscattered signal. In contrast, L-band frequencies, less sensitive to vegetation, enable detection during later canopy development, therefore accuracy remains above 80% even in late growth stages. Analysis of the 2025 season is underway.

Acknowledgment: This study is funded by ASI under the Agreement N. 2023-52-HH.1-2025 (addendum MyGEO to the THETIS project) in the framework of ASI’s program “Innovation for Downstream Preparation for Science” (I4DP_SCIENCE).

References:

[1]      C. Massari et al., “A review of irrigation information retrievals from space and their utility for users,”, Remote Sensing, 2021.

[2]      C. Corbari et al., “Estimates of Irrigation Water Volume by Assimilation of Satellite Land Surface Temperature or Soil Moisture Into a Water-Energy Balance Model in Morocco,” Water Resour Res, 61, 7, 2025.

[3]      A. Balenzano et al., “Sentinel-1 and Sentinel-2 Data to Detect Irrigation Events: Riaza Irrigation District (Spain) Case Study,” Water, 14, 19, 2022.

[4]      A. Balenzano et al., “Sentinel-1 soil moisture at 1 km resolution: a validation study,” Remote Sens Environ, 263, 2021.