Invertion of cultivated soil organic matter content combining multi-spectral remote sensing and random forest algorithm

Soil organic matter (SOM) inversion based on remote sensing generally uses single-date images as input. In order to explore the possibility of multi-spectral remote sensing with random forest to improve the accuracy of SOM inversion, this study was carried out in the cultivated land of Shengli Farm in Heilongjiang Province (133°34′-134°09′E, 47°13′-47°32′N). The Sentinel-2A and Landsat 8 images from the bare soil period were chosen as the main data sources, and were used for calculating spectral index. Random forest algorithm was used to select spectral bands and spectral index as the input variables and thus to build SOM inversion model. Results showed that: 1) the SOM spectral response band for both Sentinel-2A and Landsat 8 included the central wavelength: about 560, 660, 850 nm, and additional 740 nm of Sentinel-2A; 2) the performance of the optimal SOM inversion model, using predictors of the optimal band and spectral index in the single date from Sentinel-2A image, was well with the R2 of 0.913 and RMSEval (root mean square error for validation data) of 0.860 kg/kg, which presented better results on accuracy and stability than that of Landsat 8 image; 3) the SOM inversion accuracies using the spectral indices from Sentinel-2A and Landsat 8 images were increased by 28.87% and 8.72%, respectively compared to that using the optimal bands as input; 4) the accuracies of the inversion model based on single and double-dates bands and the spectral indices were as following: double-date images (R2 was 0.938, RMSEval was 1.329 kg/kg), Sentinel-2A image (R2 was 0.935, RMSEval was 1.944 kg/kg), Landsat 8 image (R2 was 0.922, RMSEval was 2.022 kg/kg). The stability and accuracy of the SOM optimal inversion model for double-date images was higher than that for single-date image. Red-edge band of Sentinel-2A image provided the optimal band information for the SOM inversion because its wavelength range was within the spectral response wavelength range of SOM, which was beneficial to enhance inversion accuracy. In conclusion, by applying random forest algorithm and remote sensing data and introducing spectral indices into the input, the SOM inversion accuracy could be improved and the predicted SOM map could better characterize the spatial distribution of SOM content. The results of this study proved the advantages of Sentinel-2A images and multi-temporal images in the bare soil period for SOM inversion, and can provide effective methods for improving the precision of remote sensing inversion model of soil physical and chemical parameters such as SOM.