Development of a Deep Learning Surrogate Model to Simulate MetOp ASCAT Observations with Land Surface Parameters

Abstract

In this study, deep neural networks are employed to act as a surrogate model between the Meteo France land surface model and Advanced SCATterometer (ASCAT) satellite observations. This provides a measurement operator for the assimilation of ASCAT satellite derivations into this model. Currently, TU Wien uses ASCAT measurements to retrieve soil moisture from backscatter. Next to backscatter signal, two additional vegetation parameters are extracted from the TUWien SoilMoisture Retrieval Approach. These parameters are slope and curvature and describe the second order Taylor polynomial, which explains the incidence dependency of backscatter. A recent study showed that slope and curvature could contain valuable information about vegetation water dynamics and biomass phenology. The new explored relationship gives an unique opportunity to relate land surface variables with these observation parameters. This is significant as it could be used to create a climatological data set of high quality and temporal consistency. The surrogate model avoids the need to use a Radiative TransferModel (RTM) to relate the land surface model to the ASCAT observations. RTM’s require complex input variables such as size, shape, height, thickness and orientation of the canopy but also the dielectric properties. Additionally, RTM’s are not based on the actual output of the land surface model (LSM), as the LSMs simulate vegetation parameters such as leaf area index, soilmoisture, gross primary production, temperature and respiration. Thismakes RTMs less suitable to act as a measurement operator. The suggested method to simulate ASCAT observations are deep neural networks. Deep neural networks are able to capture every highly non-linear relationship by using only the outputs from the LSMs. For this study a regular feed forward deep neural network is used to simulate the backscatter signal of the ASCAT instrument, whereas slope and curvature are simulated by a deep convolutional neural network. The results show that the deep neural network is able to simulate the seasonal and inter-seasonal variation of backscatter. Concerning slope the model was capable to capture the seasonal trends and some of the interseasonal variations. Curvature shows the worst model performance, as the model is only able to capture the timing of the seasonal changes but not the right magnitudes. In general, the performance depends on the variation of the observation and land surface data. This suggests that the model structure needs to be adapted according to the complexity of the investigated grid point. A black-box interpretation model, called DeepSHAP, is used to extract the most important features for each observation simulation. This feature importance allows a physical interpretation of the ASCAT observations. The most relevant feature for backscatter is soilmoisture, which is consistent with previous research and gives confidence to the feature importance extraction method. The slope signal is mostly related to the gross primary production and therefore biomass assimilation. Curvature shows the highest correlation for LAI. Both results confirm the previous assumptions that curvature and slope are, respectively, related to structural and phenology changes. The results of this research are substantial as they allow to the first time to relate actual vegetation parameters to the slope and curvature signals. It additionally proves that deep neural networks are a possible choice to act as a surrogate model between ASCAT observations and a land surface model.