A two-stage algorithm for regional-scale SOC prediction: Eliminating the spatial scale effect between multi-source remote sensing data

Abstract

Mapping the spatial distribution of soil organic carbon (SOC) content with high spatial resolution is important for sustainable soil management. Remote sensing variables (RSV) and environmental covariates (ECs) have been widely used for SOC content prediction and mapping. The spatial resolution mismatch between RSV and ECs, where ECs typically have lower resolution, introduces a scale effect. For example, as RSV values vary significantly within a 1 km * 1 km range, incorporating ECs at a 1 km * 1 km resolution may smooth pixel values in the mapping results, thereby compromising prediction accuracy and the mapping results. Currently, few studies have focused on how scale effects affect mapping results. Therefore, the aim of this study is to construct a two-stage algorithm for regional-scale SOC content prediction, which can effectively address the issue of spatial scale inconsistency among remote sensing images and ECs, and then combine this with convolutional neural network (CNN) model to establish a high accuracy and robust SOC content prediction model. A total of 11,841 cloudless Landsat Operational Land Imager images from 2016 to 2021 were obtained to calculate RSV, and climate, soil properties, and terrain variables were employed to reflect ECs. In the two-stage algorithm, the coefficient of variation (CV) of RSV at the spatial resolution of ECs was calculated and divided into three levels of high, medium, and low. Then we determined the level of CV of RSV, at which ECs cannot be added as input variables for SOC content prediction, thus avoiding the scale effects. The results indicated that the prediction accuracy based on RSV and ECs was higher than that based solely on RSV or ECs, while the prediction accuracy based on RSV was higher than that based on ECs. The prediction results of RSV combined with ECs obtained the best performance without considering the variability, with mean RMSE, R², RPIQ, and MAE values of approximately 5.60 g kg⁻¹, 0.74, 1.75, and 4.12 g kg⁻¹, respectively. When the variability was considered, with mean RMSE, R², RPIQ, and MAE values of approximately 4.96 g kg⁻¹, 0.78, 2.05, and 3.72 g kg⁻¹, respectively. It was found that only RSV should be used as input variables in regions with CV> 1, whereas both RSV and ECs should be used in regions with CV< 1. This could improve the prediction accuracy and eliminate the influence of the spatial scale effect of ECs on the mapping results. When the resolution difference between the RSV and ECs is 16 times or more, the influence of the spatial scale effect on the mapping results cannot be ignored. The approach proposed in this study can be applied to any problems of regression using multi-source remote sensing data of different spatial resolution, thus avoiding scale effects on the mapping results.