Quantification of high-resolution surface soil organic carbon in croplands using airborne imaging spectroscopy, soil database, and deep learning

Abstract:

Soil organic carbon content (SOC) is an essential variable to determine soil health and agricultural production. Accurate, rapid, and operational assessment of high-resolution SOC is vital for adaptive soil management towards regenerative agriculture. Traditional methods to measure SOC through field soil sampling and laboratory chemical analysis are accurate, but expensive and not scalable. Airborne imaging spectroscopy along with machine learning is a promising tool to rapidly collect high spatial and spectral resolution of soil reflectance to quantify SOC, but requires a large size of labeling dataset for modeling training. In addition, the mixture of spectral signals from bare soil, crop residue, and other materials in fields brings challenges for airborne based SOC prediction. Towards the cost-effective prediction of field-scale surface SOC, this study developed a two-step-training deep learning framework to exploit airborne imaging spectroscopy and open soil spectral libraries. In the first step, a deep neural network to predict SOC was developed using the USDA Rapid Carbon Assessment (RaCA) soil database. Based on the first-step model structure, the second step incorporated a convoluted neural network and additional layers for spectral unmixing of airborne spectra and bridge signatures of the airborne and laboratory soil spectra. To test the proposed framework, airborne hyperspectral surveys and in-situ soil sample collection were conducted at Champaign County, IL during the non-growing season of 2020. Results show that the two-step-training deep learning framework can exploit the airborne imaging spectroscopy and existing soil database to achieve high accuracy to map high-resolution SOC. Our proposed approach has a high potential for the rapid assessment of high-resolution surface SOC in croplands with limited dependence on extensive soil sampling and laboratory chemical analysis. These high-resolution surface SOC maps can provide valuable insights into soil health and facilitate the accurate quantification of the carbon budget for agroecosystems.