Integrated remote sensing and machine learning system for high-resolution soil moisture mapping

Machine-learning blends radar satellite data with one soil sample to separate soil roughness, vegetation, and moisture signals, generating 5–10 m resolution soil moisture maps adaptable to any crop or terrain worldwide.

Background

Agricultural management, flood forecasting, and drought monitoring all rely on timely, accurate information about how much water is held in the soil. Soil moisture controls plant growth, influences irrigation scheduling, and affects the reliability of water supplies. Traditional methods involve point‐based sensors or manual sampling, which provide highly accurate readings at specific locations but cannot capture the variability across an entire field or watershed. Satellite remote sensing offers broad coverage and regular revisits, yet optical and thermal imagery can be blocked by clouds and only measure surface conditions, while passive‐microwave backscatter data is often too coarse for farm‐scale decisions. There is a clear need for a method that combines different data types to create detailed, field‐level soil moisture maps that are both accurate and scalable around the world.

Current solutions struggle to disentangle the multiple factors that affect the signals used to infer moisture. Vegetation cover and surface roughness alter microwave returns, leading to ambiguous estimates when these influences overlap. Optical and thermal sensors only capture surface dryness and fail under cloudy skies, while dense canopies block microwave penetration. In‐field sensors require extensive deployment and recalibration for each crop type and terrain, making large‐area monitoring costly and labor‐intensive. As a result, farmers and resource managers lack the consistent, high‐resolution moisture information needed for optimal decision‐making.

Technology

This technology combines publicly available radar data from satellites with a single ground measurement to generate continuous and detailed soil moisture map for whole field. It estimates moisture levels across an entire field at 5–10 meter resolution. The adaptive design accommodates various crop types and terrain, and operates wherever the necessary satellite and in-field data exist, offering a practical workflow from data fusion to high-resolution mapping.

Unlike conventional soil moisture products that rely on coarse-scale satellite data, this approach delivers high spatial detail by merging radar signals. The machine learning component differentiates overlapping factors—roughness, vegetation and moisture—rather than treating them as a single blended signal. Estimation using just one field reading means the method can rapidly adapt to new locations. The reliance on publicly accessible data sources and a flexible algorithmic framework makes it both cost-effective and scalable, enabling consistent moisture monitoring across diverse agricultural landscapes worldwide.

Benefits

  • High-resolution (5–10 m) soil moisture mapping across entire fields
  • Reduced reliance on extensive in-field sampling through model extrapolation
  • Enhanced accuracy by decoupling effects of surface roughness, vegetation, and moisture
  • Adaptable to diverse crop types and terrain, enabling global application
  • Cost-effective, scalable solution leveraging publicly available satellite data
  • Supports optimized irrigation scheduling and improved water resource management

Applications

  • Precision irrigation scheduling
  • Variable rate irrigation
  • Crop yield forecasting
  • Agricultural insurance risk assessment
  • Drought monitoring services

Additional info

It uses active-microwave backscatter and applies a supervised classification/regression algorithm to decouple soil roughness, vegetation and moisture influences on the radar signal. After one in-field probe reading for calibration, it delivers 5–10 m resolution moisture estimates across diverse crops and terrains, scalable to any region with requisite satellite and ground data.