Modelling of Near-Surface Soil Moisture Using Machine Learning and Multi-Temporal Sentinel 1 Images in New Zealand

Istvan Hajdu, Ian J. Yule, Mohammad Hossain Dehghan-Shear

2018 · 36 citations · 14 references

Concepts

Abstract

The improved spatio-temporal resolution of the C-band Sentinel 1 Synthetic Aperture Radar (SAR) mission has increased the number of possibilities in soil moisture retrieval studies. In this paper, an ensemble learning method (Random Forest) was applied on a group of data acquired on a farm scale in New Zealand's hill country. A statistical model was fitted to predict volumetric soil moisture (θ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">v</sub> , %) from radar data, the degree of vegetation coverage represented by NDVI and a number of landscape parameters. The model performance demonstrated that the algorithm was able to capture the non-linear relationship between the ground-based and remotely sensed variables. θ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">v</sub> predictions using a time series of radar images reached an average accuracy of 3% for RMSE and 0.86 for R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> . An extended version of the proposed method has the potential to be a basis of more accurate water balance simulations, applied in the hydrologically complex pastoral hill country.

References

14