Inversion of soil organic matter based on GF-5 images under different noise reduction methods

Predicting the spatial distribution of Soil Organic Matter (SOM) through remote sensing images is an important part of precision agriculture and land resource management research. The emergence of Gaofen5 (GF-5) has facilitated quantitative soil monitoring based on hyperspectral remote sensing. As the first full spectrum hyperspectral satellite for comprehensive observations of the atmosphere and land, GF-5 has a high spatial resolution of 30 m, including a total of 330 bands, and the coverage is consistent with the range obtained by spectro-radiometerFieldSpec3 (Analytical Spectral Devices, Longmont, USA). However, there is no clear and appropriate noise reduction method in the relevant soil research. Therefore, this study was to explore the best noise reduction method for GF-5 application in soil organic matter inversion. The experimental area was Mingshui County (12418-12521E，4644-4729N), a typical black soil region of Heilongjiang Province. A total of 38 sampling points were selected at the depth of 0-20 cm in the study area and the spatial coordinates of sampling points were recorded by GPS. The non-cloud GF-5 hyperspectral remote sensing image was obtained from the website of China Resources Satellite Application Center. According to our prior knowledge, radiometric calibration and atmospheric correction pre-processing of hyperspectral remote sensing images in ENVI5.1 were performed. Singular Value Decomposition (SVD), Discrete Wavelet Transformation (DWT) and Median Filtering (MF) noise reduction method were used. The two-dimensional spectral indexes were calculated under different noise reduction methods. The Principal Component Analysis (PCA) method was used to decrease the dimensionality of the inputs, and the Random Forest (RF) method was used to predict the SOM content. A total of 26 data were used for model establishment and the rest 12 data were used for model validation. The SOM content of these samples were 3.45%-5.53% with mean of 4.44%, standard deviation of 4.28% and coefficient of variance of 9.6. The results showed that the best decomposition layer of DWT was 1 since the correlation coefficient between reflectance and SOM was the highest and the wavelength locations with high correlation were consistent with those in the original reflectance curves. Among different noise reduction methods, the correlation between the spectral curve and the SOM content was the highest for DWT, followed by SVD and MF. The small "burr" in the soil reflection spectrum curve was removed by using different noise reduction methods and the shapes of the spectral curves under different SOM contents were consistent. The principal component analysis results showed that the cumulative contribution rates for data experiencing noise reduction by SVD, DWT and MF methods were all about 95%. The DWT method combined with RF prediction model achieved the highest accuracy of SOM estimation for validation dataset (R2: 0.69, RMSE: 2.26%, RPD: 1.80) and the noise reduction effect was the most significant. The SOM spatial distribution under the DWT noise reduction model was mapped and the result was consistent with the real condition. The SOM content was relatively high and its distribution was uniform in the central areas of Mingshui County. However, it was low in the southeast region because of severe soil erosion. The research results can provide an effective way of the realization of remote sensing prediction of SOM at a large scale by using GF-5 hyperspectral satellite data, expanding the application scope of satellite images.