Mapping Soil Organic Carbon Stock by Hyperspectral and Time-Series Multispectral Remote Sensing Images in Low-Relief Agricultural Areas
Long Guo,Xiaoru Sun,Peng Fu,Tiezhu Shi,Lina Dang,Yiyun Chen,M. Linderman,Ganlin Zhang,Yu Zhang,Qinghu Jiang,Haitao Zhang,Chen Zeng
DOI: https://doi.org/10.1016/j.geoderma.2021.115118
IF: 6.1
2021-01-01
Geoderma
Abstract:High-precision digital soil organic carbon (SOC) stocks mapping is very important for agricultural production management and global carbon cycle. The spatial heterogeneity of farmland SOC is not only influence by the environmental factors of soil formation but also the management practices of tillage, fertilization, and irrigation. However, the traditional modeling covariates of digital soil mapping, such as terrain factors, land use types and climate factors have weak spatial variations in low-relief agricultural areas, and they cannot reflect the large spatial variation of SOC. Thus the time-series multispectral remote sensing images will be used for mapping soil properties in low relief regions in this study, meanwhile a new collaborative verification strategy was put forward to evaluate the spatial distribution characteristics of soil maps. The current study was performed in a nearly flat agricultural region southeast of Iowa (with an area of approximately 385.45 ha), where 195 surface soil samples (0-15 cm) were collected. A hyperspectral image (Headwall-Hyperspec, 380-1700 nm) and the time-series multispectral remote sensing images of Sentinel 2 and Landsat 8 were used to construct the prediction models of SOC stock and its relevant soil properties of SOC and soil bulk density (SBD) through partial least square regression (PLSR) and extreme learning machine (ELM) models. The collected soil samples and evaluation indexes of root mean square error (RMSE), R-2, and ratio of performance to interquartile range (RPIQ) were used to evaluate the model performance. Results are as follows: (1) hyperspectral images were successfully used to predict the SOC stock, SOC, and SBD through PLSR and ELM, while ELM (RPIQ = 2.03, 1.97, 1.64) outperformed PLSR (RPIQ = 1.83, 1.97, 1.53); (2) the time-series multispectral remote sensing images of Sentinel 2 and Landsat 8 can reflect the spatial distribution characteristics of the SOC stock, SOC and SBD by PLSR and ELM, but the combination of Sentinel 2 images and ELM obtained the best prediction results (RPIQ = 1.45, 1.25, 1.26); and (3) the differences of the soil maps predicted by the hyperspectral image and time-series multispectral remote sensing images were small, and the largest percentage errors nearly appeared on the edges of the farmland patches owing to mixed pixels. This study further confirmed the good prediction abilities of the time-series multispectral remote sensing images in low relief farmland regions. Lastly, this mapping strategy can provide additional valuable information for agricultural management and carbon cycle.