An advanced soil organic carbon content prediction model via fused temporal-spatial-spectral (TSS) information based on machine learning and deep learning algorithms

Knowledge of the soil organic carbon (SOC) content is critical for environmental sustainability and carbon neutrality. With the development of remote sensing data and prediction models, the comprehensive utilization of multisource remote sensing data based on a fusion approach and testing its effectiveness in SOC content prediction is an interesting and challenging topic. However, there is no evidence showing the role of different data sources in the SOC content prediction process. In this study, a total of 796 topsoil samples (0–20 cm) were collected at Site 1, and 111 samples were collected at Site 2. The samples from Site 2 were used to verify the transferability of the prediction model established at Site 1. The discrete wavelet transform based on the regional energy weight (RW-DWT) and spectral band segmentation methods were used to fuse the temporal information of 10 scenes of Landsat multispectral image data from 2009 to 2019, the spatial information of topography data and the spectral information of GaoFen-5 hyperspectral images. Then, the SOC content prediction models were established by temporal-spatial-spectral (TSS) information using partial least squares regression (PLSR), random forest (RF) and convolutional neural network (CNN) algorithms. The results indicated that the optimal SOC content prediction model consisted of TSS information as input and the CNN as the prediction model, where the lowest root mean square error (RMSE) was 2.49 g kg −1 , the highest coefficient of determination (R 2 ) was 0.86 and the ratio of performance to interquartile distance (RPIQ) was 1.91. Next, the order of the effect was spectral > temporal > spatial information in terms of SOC content prediction, and their roles in improving the accuracy of the model were 26.79%, 19.64% and 14.29%, respectively, with the CNN model. In addition, the CNN yielded a higher prediction accuracy than PLSR and RF regardless of which group of input variables was used. The average RMSE of the CNN was 0.42 g kg −1 lower than that of the RF, and the average R 2 and RPIQ were 9.25% and 0.14 higher, respectively, than those of the RF. The above conclusions were confirmed in the verification area, namely, the optimal SOC content prediction model at Site 2 consisted of TSS information as input and the CNN as the prediction model (RMSE = 1.01 g kg −1 , R 2 = 0.76 and RPIQ = 1.41). Therefore, the novel method proposed in this study is robust. This work provides a new idea for predicting soil properties by the comprehensive use of multisource remote sensing images and deep learning algorithms in the future. • The ability of TSS information in SOC prediction is determined. • The SOC prediction model with high-accuracy and high-transferability is established. • The method of temporal, spatial, spectral information fusion is improved. • The role of temporal, spatial, spectral information for SOC prediction is revealed. • “Data fusion + deep learning” strategy provide a new paradigm for SOC prediction.