Estimating Soil Organic Matter Content Using Sentinel-2 Imagery by Machine Learning in Shanghai

Soil organic matter (SOM) plays an important role in the field of climate change and terrestrial ecosystems. SOM in large areas, especially in urban areas, is difficult to monitor and estimate by traditional methods. Urban land structure is complex, and soil is a mixture of organic and inorganic constituents with different physical and chemical properties. Previous studies showed that remote sensing techniques that provide diverse data in the visible-near-infrared (VNIR)-shortwave infrared (SWIR) spectral range, are promising in the prediction of SOM content on a large scale. Sentinel-2 covers the important spectral bands (VNIR-SWIR) for SOM prediction with a short revisit time. Thus, this article aimed to evaluate the capacity of Sentinel-2 for SOM prediction in an urban area (i.e., Shanghai). 103 bare soil samples filtrated from 398 soil samples at a depth of 20 cm were selected. Three methods, partial least square regression (PLSR), artificial neural network (ANN), and support vector machine (SVM), were applied. The root mean square error (RMSE) of modelling (mRMSE) and the coefficient of determination (R-2) of modelling (mR(2)) were used to reflect the accuracy of the model. The results show that PLSR has the poorest performance. ANN has the highest modelling accuracy (mRMSE = 7.387 g kg(-1), mR(2) = 0.446). The ANN prediction accuracy of RMSE (pRMSE) is 4.713 g kg(-1) and the prediction accuracy of R-2 (pR(2)) is 0.723. For SVR, the pRMSE is 4.638 g kg(-1), and the pR(2) is 0.732. The prediction accuracy of SVR is slightly higher than that of ANN. The spatial distribution of SOM demonstrates that the value obtained by ANN is the closest to the range of the bare soil samples, and ANN performs better in vegetation-covered areas. Therefore, Sentinel-2 can be used to estimate SOM content in urban areas, and ANN is a promising method for SOM estimation.