Predicting spatial–temporal soil organic matter dynamics in a Mollisols region of the northern Songnen Plain, China, during 2009–2018 using a spectral-temporal feature set

Similar to many other parts of the world, it is a necessity to reveal the spatial–temporal soil organic matter (SOM) dynamics of Mollisols in the northern Songnen Plain, a state key agricultural region in China. Although digital soil mapping (DSM) with temporal environmental covariates could fulfill this purpose, its accuracy still needs to be improved. The present study aimed to evaluate whether a spectral-temporal feature set derived from percentile transformations of time-series remote sensing images could be helpful for improving the accuracy, because the feature set is advantageous in providing wall-to-wall and stable information and having large dimensionality. The evaluation was conducted in the case of the Mollisols region, where a total of 334 soil samples were collected during 2009–2011 and 2014–2018 and measured for SOM contents. As environmental covariates, a spectral-temporal feature set consisting of a series of percentiles (i.e., 10 %, 25 %, 50 %, 75 %, and 90 %) of spectral bands and indices and corresponding means were derived from the MODIS/Terra images for every five years between 2009 and 2018, while the terrain and climate factors were also obtained. With these data, classification and regression tree (CART) and random forest (RF) were both employed to establish spatiotemporal models for predicting SOM content at five-year intervals from 2009 to 2018. Results showed that replacing the commonly used means and medians of spectral bands and indices in the two machine learning models with the spectral-temporal feature set improved the prediction accuracy, with an increase of the mean concordance correlation coefficient (CCC) by 1.94 %∼8.09 %. Further, the optimal RF model with the spectral-temporal feature set was used to generate SOM content maps for every five years between 2009 and 2018, which were validated based on the samples of 2009–2011, showing a CCC of 0.66. The resulted maps showed that the mean SOM content decreased from 2009 to 2018 by 0.04 %. An importance analysis showed that lots of the spectral-temporal features were the most important variables in the RF model, following the first important one (i.e., mean annual precipitation). As there were many spectral bands and indices, their sum importance was far larger than all the other kinds of environmental covariates. It is concluded that the spectral-temporal feature set is promising for deriving the spatial–temporal dynamics of soil in the future.