An advanced soil organic carbon content prediction model via

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

In the global carbon cycle, soil organic carbon (SOC) is the largest terrestrial carbon reservoir, which accounts for approximately 50-80% of the total terrestrial carbon, contains more than three times as much as the atmosphere or vegetation, and determines a landscape's carbon source/sink ability. Knowledge of the 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.

For this, a group of Chinese scientists from Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences collected soil samples from a cultivated land of the Mollisol region of Northeast China (NEC) (123°44′-126°41′ E, 46°38′-48°07′ N) (site 1) and Aohan Banner, the southeastern Inner Mongolia Autonomous Region, China (41°69′-43°03′ E, 119°53′-120°89′ N) (site 2). They measured the SOC content and in-situ topsoil spectral data (ASD FieldSpec 3 Spectroradiometer), and acquired GF-5 satellite hyperspectral image and Landsat image. 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 specific objectives of this study are (1) to develop a new method to extract TSS fusion information and validate whether TSS information could better represent soil information and improve the accuracy of SOC content prediction; (2) to develop an advanced SOC content prediction model with TSS information and machine learning and deep learning algorithms, compare them with temporal, spectral, spatial, temporal-spectral, temporal-spatial or spatial-spectral information as input, and verify the transferability of the model; and (3) to evaluate the role of temporal, spatial and spectral information for predicting the SOC content.

Overview of study sites. (a) the middle of Northeast China (Site 1), the location of in-situ spectral sampling points and topsoil sampling points; (b) the soil classes and the cultivated land range of Site 1; (c) the Aohan Banner (Site 2) and the location of topsoil sampling points; (d) the soil classes and the cultivated land range of Site 2; (e) photographs of the soil surface during the bare soil period; (f) collected topsoil sampling points within a quadrat.

Flowchart of the prediction of SOC content from multi-temporal multispectral, hyperspectral, topography, and their fusion data.

Flowchart of the prediction of SOC content from multi-temporal multispectral, hyperspectral, topography, and their fusion data.

Results：

Image fusion process based on the regional energy weighting discrete wavelet transform. (a) the process of fusing multi-temporal multispectral image; (b) the hyperspectral reflectance curves of in-site soil samples; (c) the process of fusing simulated and real hyperspectral image.

Scatter plot between lab-measured and predicted SOC for different input variables, using the CNN prediction model at Site 1 and Site 2.

Conclusions: 

(1) a new research paradigm for SOC content prediction was developed, namely, “data fusion + deep learning”. When TSS information (FHI + TI) and CNN were used as input and prediction models, the highest SOC content prediction accuracy was obtained because TSS information contains advantageous information of multi-source remote sensing data;

(2) the RW-DWT fusion method effectively obtained comprehensive information of temporal, spatial, and spectral information to improve the representation of soil information and the accuracy of SOC content prediction;

(3) the CNN performed better than PLSR and RF in SOC content prediction because it could use the spatial contextual information and avoided the simple relationship that only considers an SOC sample and a single pixel in PLSR and RF;

(4) although the SNR of GF-5 may be low in the SWIR region, the prediction accuracy of the hyperspectral sensor was significantly better than that of the Landsat-8 OLI sensor because numerous SOC-sensitive bands and narrower bandwidths enable better exploitation of the spectral characteristics corresponding to the SOC content;

(5) the temporal information reduced the negative impact of soil moisture on SOC content prediction accuracy, but the accuracy was not proportional to the amount of temporal information. When MMI7* was used as input, the highest prediction accuracy and the lowest NDMI value were observed;

(6) the order of different types of information in the improvement of SOC content prediction accuracy was spectral > temporal > spatial. The role of this information was understood, which will help us understand which information is more conducive to SOC content prediction. This work provides a new research paradigm for soil property parameter prediction in other Mollisol areas at the same latitude. Future research should explore more efficient fusion methods and deep learning algorithms.