Use of advanced machine learning for improved exploitation of remote sensing information
Sujay Kumar(1), Shahryar Ahmad(1), Goutam Konapala(1), Clara Draper(2)
(1) NASA GSFC
Abstract: Given the significant heterogeneity and complexity of the land surface, there are significant barriers to fully exploiting the information content of remote sensing datasets. While there has been significant progress in the use of land data assimilation methods, majority of them still rely on the use of retrieval products to incorporate them within land surface models. The reliance on retrieval model products, which have their own associated biases and uncertainties has been limiting. In soil moisture data assimilation instances, for example, remote sensing retrievals are often rescaled to match the climatology of the model because of the large scale systematic differences between the model and remote sensing retrieval estimates. These rescaling approaches lead to loss of information and are inadequate in handling dynamic changes in bias characteristics, and when unmodeled processes are present. The use of optimization tools to reduce the systematic errors in the model, therefore, is desirable. The traditional calibration approaches, however, are computationally expensive, limiting their application over large/fine spatial scales. Here we demonstrate the use of advanced machine learning tools for the effective reduction of systematic errors in a computationally efficient manner. The presentation will also discuss how the use of machine learning is impactful in improving the information content of retrieval products. For example, though there has been a long legacy of passive microwave radiometry for snow mass estimation, most of the retrievals are fraught with issues of limited skill over mountains and forests and insufficient interannual variability. The use of advanced machine learning tools is more effective in exploiting the relative sensitivities in radiance measurements for improving these limitations. The presentation will also describe how the machine learning applications provide inferences on improving model representations.
Assimilating ASCAT dynamic vegetation parameters to constrain the plant water dynamics in land surface model
Xu Shan(1), Susan Steele-Dunne(1), Manuel Huber(2), Sebastian Hahn(1), Wolfgang Wagner(1,3), Bertrand Bonan(4), Clement Albergel(4), Jean-Christophe Calvet(4), Ou Ku(5), Sonja Georgievska(5)
(1) Department of Geoscience and Remote Sensing, Faculty of Civil Engineering and Geosciences, TU Delft, Delft, the Netherlands
(2) Department of Water Management, Faculty of Civil Engineering and Geosciences, TU Delft, Delft, the Netherlands; now at European Space Agency, European Space Research and Technology Centre (ESTEC), 2201 AZ, Noordwijk, the Netherlands
(3) Department of Geodesy and Geoinformation (GEO), Vienna University of Technology, Vienna, Austria
(4) CNRM, Université de Toulouse, Météo-France, CNRS, Toulouse, France; now at European Space Agency Climate Office, ECSAT, Harwell Campus, Didcot, Oxfordshire, UK;
(5) Netherlands eScience Center, Amsterdam, the Netherlands
Abstract: Our current ability to parameterize plant water dynamics in land surface model (LSM) constrains our capacity to understand land-atmosphere processes, and our ability to represent the response of ecosystems to drought (Powell et al., 2013). Microwave remote sensing datasets contain valuable information about plant water content variations from sub-daily scale to interannual scales under saturation or water stress (Konings et al., 2017; Steele-Dunne et al., 2019) and can be assimilated to constrain plant water dynamics in LSMs.
Recent research has shown that the backscatter-incidence angle relationship of the Advanced Scatterometer (ASCAT) data varies in response to changes in vegetation water content and phenology. In this study, we are working towards assimilating these data to constrain water dynamics in a LSM. Firstly, we addressed the challenge about how to reconcile the states and parameters of the LSM with the satellite observations. A Deep Neural Network (DNN) was trained to link the ASCAT observables to the soil moisture in different layers and the vegetation-related states. Secondly, we assimilated ASCAT dynamic vegetation parameters into land surface model ISBA-A-gs.
In a study over France from 2007 to 2019, the DNN is used to simulate the normalized backscatter as well as the slope and curvature of the backscatter-incidence angle relationship. Results show that the DNN has a near zero bias for normalized backscatter and slope. A sensitivity analysis shows that ASCAT observables are sensitive to variations in not only surface soil moisture and LAI, but also root zone soil moisture because of the dependency of plant water content on soil moisture in deeper layers.
Further, assimilation results show improvement of estimates of soil moisture and LAI. Furthermore, this method is highly transferable and lends itself to multi-observation assimilation. This paves the way to constrain vegetation water processes in LSMs using all available satellite data.
Contribution of machine learning for the integration of satellite observations in a global model of the soil-plant system
Timothée Corchia(1), Bertrand Bonan(1), Jean-Christophe Calvet(1), Gabriel Colas(1), Nemesio Rodriguez-Fernandez(2)
(1) CNRM, Université de Toulouse, Météo-France, CNRS, 31057, Toulouse, France
(2) Centre d’Etudes Spatiales de la Biosphère, CESBIO – CNESCNRSIRSUPS
Abstract: In the context of climate warming, the frequency and the intensity of extreme events such as droughts is increasing and better modeling of the response of vegetation to climate is needed. Monitoring the impact of extreme events on terrestrial surfaces involves a number of variables of the soil-plant system such as surface albedo, the soil water content and the vegetation leaf area index (LAI). These variables can be monitored by either using the unprecedented amount of data from the Earth observation satellite fleet, or using land surface models. Another solution consists in combining all available sources of information by assimilating satellite observations into models. In this work, C-band Advanced SCATterometer (ASCAT) Radar backscatter (sigma0), L-band Soil Moisture and Ocean Salinity (SMOS) Vertical and Horizontal brightness Temperatures (V and H BT) satellite products are assimilated in the ISBA land surface model of Meteo-France using the LDAS-Monde tool. First, observation operators are built using machine learning. Neural networks (NNs) are trained using the modeled surface soil moisture (SSM), soil temperature, rainwater interception by leaves, and satellite-derived LAI observations from Copernicus as inputs. The NNs are then used to find the statistical relationship between the input data and the satellite products, making LDAS-Monde capable of assimilating the satellite observations. It is shown that the assimilation of level 1 data alone is able to markedly improve the simulated LAI and SSM.
PROcess-guided deep learning and DAta-driven modelling (PRODA) to uncover key patterns and mechanisms in global soil carbon dynamics
Feng Tao(1), Yiqi Luo(2)
(1) Department of Earth System Science, Tsinghua University, Beijing, 100084, China
(2) Center for Ecosystem Science and Society, Department of Biological Sciences, Northern Arizona University, AZ, USA
Abstract: Soils are the largest organic carbon pool in the terrestrial ecosystem. Yet, key mechanisms that regulate soil organic carbon (SOC) formation and sequestration remain poorly understood. To better understand global SOC storage and its feedback to changing climate, we developed a novel PROcess-guided deep learning and DAta-driven modelling (PRODA) approach. PRODA integrates data assimilation, deep learning, big soil carbon datasets, and process-oriented models to best represent and understand global soil carbon dynamics. In an example that integrated 52,819 globally distributed vertical SOC profiles into the Community Land Model (CLM5), PRODA-optimised model simulation explained 57% spatial variation in SOC content. Meanwhile, microbial carbon use efficiency (CUE) emerged as the pivot to global SOC storage and spatial distributions compared with other mechanisms (e.g., decomposition, plant carbon input, and vertical transport). The findings revealed by PRODA enriched the classic paradigm to focus on not only SOC decomposition and organic carbon input but also microbial CUE in understanding global SOC formation and persistence. Moreover, PRODA approach presents its potential in gaining emergent understandings of transient dynamics of SOC from integrating multiple sources of soil carbon datasets into process models. In an example at Harvard Forest, SOC sequestration from 1900 to 2010 after being constrained by both SOC content and soil radiocarbon data showed higher efficiency with lower residence time than the results only informed by the radiocarbon data. In the future, integrating process-oriented models with different sources of global soil carbon datasets is essential to accurately quantify global soil sequestration under climate change.
Comparative evaluation of different data assimilation approaches to optimize the parameters of the ORCHIDEE land surface model
Philippe Peylin(1), Nina Raoult(1), Maxime Carenso(1), Vladislav Bastrikov(1), Catherine Ottle(1), Maelle Coulon(1), James Salter(1), Cedric Bacour(1) and the ORCHIDAS group
(1) Laboratoire des Sciences du Climat et de l’Environnement (LSCE)
Abstract: For more than 10 years, different approaches have been developed by the international scientific community to optimize the parameters of biosphere models by assimilating different types of observations. These are essentially based on a Bayesian formalism with the minimisation of a cost function that takes into account all the errors associated with the model (structural errors and errors associated with the parameters) as well as the observations and our a priori knowledge of the parameters (assuming also Gaussian error distributions). Key examples include variational approaches (i.e., using a gradient method which requires the calculation of the sensitivity of the cost function to the parameters), Monte Carlo approaches (genetic algorithms, Markov chains, etc.) or “filter” approaches (i.e. Kalman filter, particle filter, etc.). Within the framework of the optimisation of the global continental surface model, ORCHIDEE, we have developed an assimilation system (ORCHIDAS) and tested mainly 2 methods (gradient method and genetic algorithm): see https://orchidas.lsce.ipsl.fr/. However, recent developments have highlighted alternative methods, based on ensemble filters or using physical model emulators, which offer advantages, particularly with regard to i) the numerical speed of the optimisation and ii) the ease of assimilating a set of observations of various natures. In this presentation, we look in particular at History Matching – a method based on emulation techniques developed by the uncertainty quantification community for the calibration of model parameters and successfully applied to climate models. We discuss the advantage of this technique (with respect to the gradient method and a genetic algorithm) and test the potential of this new approach in calibrating ORCHIDEE. The test case will consist in assimilating in situ data of water and carbon flux measurements (about 100 sites) and satellite proxies of vegetation activity (solar-induced fluorescence – SIF) to evaluate the respective performances of the different methods: level of fit to the observations, sensitivity to local minima of the cost function, computation time, etc.
Optimizing rain gauge locations based on data-driven sparse sensor placement
Daiya Shiojiri(1), Takumi Saito(1), Mao Ouyang(1), Shunji Kotsuki(1)
(1) Center for Environmental Remote Sensing, Chiba University
Abstract: Precipitation is one of the most important variables in hydrological studies because it provides the ultimate source of water resources, and occasionally causes severe disasters. Therefore, estimating spatio-temporal distributions of precipitation has been a key challenge in hydrological studies. Rain gauge stations provide essential ground truth data that is used to calibrate satellite/ground radar observations and numerical weather prediction models. However, there have been few studies that have explored optimization methods of rain gauge locations. For the rain gauge placement in this study, we use the data-driven sparse sensor placement (SSP) method, which has been developed in informatics science. This method determines the optimal sensor locations so that the selected sensors effectively determine coefficients of proper orthogonal decomposition (POD) modes. The original SSP method reconstructs the spatial patterns of data from the selected sensors by solving a linear inverse problem using the POD modes.
This study extends the existing SSP method for the problem of rain gauge placements by incorporating singular values of POD in addition to the POD modes. We also introduce two to reconstruct spatial patterns of precipitation. One is the data assimilation approach that can estimate the spatial patterns better than the simple linear inverse problem owing to Tikhonov regularization. The other implementation is the localization for eliminating erroneous sampling noise and increasing the rank of background error covariance. We applied the proposed method for the placements of rain gauge observations over Hokkaido Island in Japan. Here we used 14-day accumulated precipitation of radar observations from 2006 to 2016 as training data. Based on the POD modes of the training data, observation locations were determined. We estimate spatial patterns of precipitation from selected points of precipitation by data assimilation, and compared with reference radar data for 2017-2018. The optimized locations of rain gauge stations by SSP method reconstruct more accurate spatial patterns of precipitation than the fields reconstructed with operationally distributed rain gauge locations.