Soil Moisture Content Inversion Model Based on Landsat8 and Sentinel-1 Image Fusion

Abstract

To address the current problems that a single optical satellite is easily affected by clouds and SAR satellite is easily affected by vegetation and soil roughness when being applied into soil moisture content inversion, taking Shahaoqu of Hetao Irrigation Area as study area, and taking soil moisture content of four depths in April 2019 as study object, PCA and GS were used to fuse Landsat8 and Sentinel-1 images and the quality of the fused images was evaluated. Then a total of 1134 remote sensing indices were constructed with the gray value of the fused images, and soil moisture content inversion models were constructed based on three variable screening methods (correlation coefficient analysis, variable projection importance analysis and gray correlation analysis) and four machine learning algorithms (BP, ELM, RF, and SVM). The study results showed that the fused images of PCA and GS fusion could successfully maintain the advantages of both Sentinel-1 and Landsat8 images in quantitatively inversion of soil moisture content. The three-dimension indices constructed based on the fused images were generally more sensitive to soil moisture content than two-dimension indices constructed based on fused images. The VIP-ELM model based on GS fusion had the highest accuracy in the surface soil moisture content inversion (R2=0.66, RMSE was 1.35%). When VIP-ELM model based on GS fusion was applied to the soil moisture content inversion at all depths, 20~40cm achieved the best performance (R2=0.79, RMSE was 0.94%), followed by 0~10cm, 40~60cm and 10~20cm. This finding can provide a strong reference for using multi-source satellite image fusion to monitor soil moisture content.

Keywords: soil moisture content, satellite image fusion, machine learning, coupled model