Extracting Soil Moisture from Fengyun-3D Medium Resolution Spectral Imager-II Imagery by Using a Deep Belief Network

+ Author Affiliations + Find other works by these authors
  • Corresponding author: Chengming ZHANG, chming@sdau.edu.cn
  • Funds:

    Supported by the Science Foundation of Shandong (ZR2017MD018); Key Research and Development Program of Ningxia (2019BEH03008); Open Research Project of the Key Laboratory for Meteorological Disaster Monitoring, Early Warning and Risk Management of Characteristic Agriculture in Arid Regions (CAMF-201701 and CAMF-201803); Arid Meteorological Science Research Fund Project by the Key Open Laboratory of Arid Climate Change and Disaster Reduction of China Metrological Administration (IAM201801); and Science Foundation of Ningxia (NZ12278)

  • doi: 10.1007/s13351-020-9191-x

PDF

  • Obtaining continuous and high-quality soil moisture (SM) data is important in scientific research and applications, especially for agriculture, meteorology, and environmental monitoring. With the continuously increasing number of artificial satellites in China, the acquisition of SM data from remote sensing images has received increasing attention. In this study, we constructed an SM inversion model by using a deep belief network (DBN) to extract SM data from Fengyun-3D (FY-3D) Medium Resolution Spectral Imager-II (MERSI-II) imagery; we named this model SM-DBN. The SM-DBN consists of two subnetworks: one for temperature and the other for SM. In the temperature subnetwork, bands 1, 2, 3, 4, 24, and 25 of the FY-3D MERSI-II imagery, which are relevant to temperature, were used as inputs while land surface temperatures (LST) obtained from ground stations were used as the expected output value when training the model. In the SM subnetwork, the input data included LSTs generated from the temperature subnetwork, normalized difference vegetation index (NDVI), and enhanced vegetation index (EVI); and the SM data obtained from ground stations were used as the expected outputs. We selected the Ningxia Hui Autonomous Region of China as the study area and used selected MERSI-II images and in-situ observation station data from 2018 to 2019 to develop our dataset. The results of the SM-DBN were validated by using in-situ SM data as a reference, and its performance was also compared with those of the linear regression (LR) and back propagation (BP) neural network models. The overall accuracy of these models was measured by using the root mean square error (RMSE) of the differences between the model results and in-situ SM observation data. The RMSE of the LR, BP neural network, and SM-DBN models were 0.101, 0.083, and 0.032, respectively. These results suggest that the SM-DBN model significantly outperformed the other two models.
  • 加载中
  • Fig. 1.  (a) Geographical location of the Ningxia Hui Autonomous Region of China from a Fenyun-3D image. (b) Distribution of sites for soil moisture observations (orange triangles) and ground meteorological stations (green pentagons) in the same area.

    Fig. 2.  Structure and procedure of SM-DBN modeling.

    Fig. 3.  The structure of each RBM.

    Fig. 4.  Distributions of SM over the Ningxia Hui Autonomous Region of China based on the (a) SM-DBN, (b) SM-LR, and (c) SM-BP models for five days: 6 March, 3 April, 23 May, 29 October, and 31 October 2019 (from the left to right panels sequentially).

    Fig. 5.  Correlation between the observed and predicted SM based on the (a) SM-DBN, (b) SM-LR, and (c) SM-BP models. The dashed line in each panel denotes the fitting line from all testing points.

    Fig. 6.  Time series of the observed and predicted soil moisture from SM-DBN over (a) Tongxin and (b) Ligang during 16–30 May and 1–15 April 2018, respectively.

    Table 1.  Training parameters for the temperature subnetwork

    Momentum0.1
    Structure[n, n × 3, n × 5, n × 7, n × 9, n × 10, n × 8, n × 6, n × 4, n × 2, 1]
    cd_k1
    RBM learning ratee–4
    RBM epoch600
    BP learning ratee–4
    Dropout0.0005
    Batch size50
    Note: structure denotes the number of RBM layers and the number of neurons of each layer; n denotes the number of input parameter; cd_k denotes the sampling times; learning rate denotes the parameter in the optimization algorithm that determines the step size at each iteration while moving toward the minimum of the loss function; RBM epoch denotes RBM training times of each layer; dropout denotes the probability of abandonment of neurons; batch size denotes the number of samples per training.
    Download: Download as CSV

    Table 2.  Training parameters for the SM subnetwork

    Momentum0.1
    Structure[n, n × 8, n × 14, n × 16, n × 17, n × 18, n × 12, n × 11, n × 10, n × 9, n × 6, n × 2, 1]
    cd_k1
    RBM learning ratee–4
    RBM epoch200
    BP learning ratee–4
    Dropout0.0005
    Batch size50
    Download: Download as CSV

    Table 3.  Number of samples used in each round of the comparison experiments

    TypeTest sampleTraining sample
    Temperature346013,840
    SM2342 9368
    Download: Download as CSV

    Table 4.  Accuracy of the SM prediction from three models

    ModelR2RMSE
    SM-DBN0.9130.032
    SM-LR0.6380.101
    SM-BP0.8130.083
    Download: Download as CSV
  • [1]

    Abu-Hamdeh, N. H., 2003: Thermal properties of soils as affected by density and water content. Biosyst. Eng., 86, 97–102. doi: 10.1016/S1537-5110(03)00112-0.
    [2]

    Adegoke, J. O., and A. M. Carleton, 2002: Relations between soil moisture and satellite vegetation indices in the U.S. Corn Belt. J. Hydrometeor., 3, 395–405. doi: 10.1175/1525-7541(2002)003<0395:RBSMAS>2.0.CO;2.