|VONG, CHIN NEE - University Of Missouri|
|Sudduth, Kenneth - Ken|
|ZHOU, JIANFENG - University Of Missouri|
Submitted to: International Conference on Precision Agriculture Abstracts & Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 5/14/2022
Publication Date: 6/26/2022
Citation: Ransom, C.J., Vong, C., Sudduth, K.A., Kitchen, N.R., Veum, K.S., Zhou, J. 2022. Estimating soil carbon stocks with in-field visible and near-infrared spectroscopy. Proceedings of the 15th International Conference on Precision Agriculture, June 26-29, 2022, Minneapolis, Minnesota. Available: https://ispag.org/proceedings/
Interpretive Summary: Sequestering carbon in the soil is one approach to carbon drawdown that may help mitigate global warming. However, measuring carbon sequestration in soil is time-consuming and expensive. Probes with sensors that measure the reflectance of the soil have the potential to make this process much more rapid and inexpensive. However, relating soil reflectance measurements to soil carbon is complex. Using deep learning algorithms could help advance and improve this technique. A preliminary test of deep learning algorithms showed no improvement over the traditional methods used to relate soil reflectance to soil carbon. However, additional techniques will be evaluated to improve the accuracy of the deep learning algorithm. Overall, this study contributes to the development and refinement of deep learning techniques for complex environmental and agronomic applications that will benefit producers, scientists, and other stakeholders.
Technical Abstract: Agricultural lands can be a sink for carbon and play an important role in offsetting carbon emissions. Current methods of measuring carbon sequestration—through repeated temporal soil samples—are costly and laborious. A promising alternative is using visible and near-infrared (VNIR) diffuse reflectance spectroscopy. However, VNIR data are complex, which requires several data processing steps and often yields inconsistent results, especially when measurements are taken in situ. Using a convolutional neural network (CNN) could bypass these steps and incorporate measurements from multiple sensors to predict three-dimensional carbon stocks. Using data previously collected (n = 1,069; from 2014 to 2020), a CNN modeling framework was developed to predict soil carbon by incorporating information from profile VNIR, soil apparent electrical conductivity (ECa), penetration resistance, and soil moisture. The developed CNN framework demonstrated moderate accuracy for predicting soil carbon (r2 =0.68) but was not as accurate as the partial least squares regression model (PLSR) used as a benchmark (R2 =0.70). Additional improvements to the CNN results could occur with further attempts to optimize the parameters.