Monitoring of soil moisture in the South of Ukraine using active and passive remote sensing data

摘要

The paper considers the application of artificial neural networks for drought monitoring using Sentinel-1 and Sentinel-2 satellites' data in the South of Ukraine. From the data collected from several weather stations equipped with suction pressure measuring Watermark sensors in 2017-2018, we formed 5 datasets containing suction pressure; NDVI and NDWI values calculated using Sentinel-2 images; digital number values of co and cross-polarized radar data obtained from the Sentinel-1 images. Four datasets were used to train a neural network, and the fifth one - for accuracy checking. We used a multilayer perceptron neural network to detect the dependencies between suction pressure and normalized values of spectral indices combined with radar data. The best accuracy within the training datasets was obtained for the neural network with one neuron in one hidden layer and one neuron in the input layer with sigmoidal transfer function (maximal relative error of 28.0% or 17 kPa in absolute values). Estimates for the testing dataset described the actual data with an average relative error of 32.0%. The neural network better estimated lower levels of moisture content that is essential while predicting soil droughts. We also compared the results of neural network assessment with several spectral drought indices. As a proof of adequacy, for three images acquired in 2018, the estimates obtained by the neural network correlated with ln(TVX) (temperature vegetation index) with R=0.44-0.62. The neural network approach with the use of Sentinel imagery, however, allows evaluating soil moisture content with greater time resolution.