Advancements in modeling soil carbon using remotely sensed data

Abstract

This research investigated the resolving of soil carbon and interconnected properties in conjunction with vegetation and landform attributes for a boreal context within the Great Clay Belt region of northern Ontario, Canada. Objectives entailing sophisticated statistical inference of extracted soil samples, an endemic tree species classification and enhanced digital soil mapping of carbon were explored, yielding research contributions and practicable results. Statistical inference was achieved by adapting generalized estimating equations (GEEs) to estimate and assess refined mean differences of soil carbon between specific land cover types and tree species groupings. Furthermore, the feasibility of GEEs for digital soil mapping was demonstrated by predicting carbon concentrations with associated confidence intervals. Regarding boreal vegetation modeling, a pixel-based tree species classification was developed for natural settings at the remarkable spatial resolution of 2 m over an area of 100 square km. Afterwards, digital soil mapping of carbon was implemented by employing data-driven approaches via encoder-decoder (ED) neural networks formulated for relatively smaller data sets. Using EDs, modeling accuracy for soil carbon was augmented with coefficients of determination (R-squared) increasing to over 0.5. Finally, knowledge-based approaches to unravel linkages between soil carbon and interrelated properties were accommodated through structural equation modeling (SEM). A framework was devised to effectively facilitate SEM for digital soil mapping. The SEM quantified impacts on soil properties from covariates relating to soil formation factors, supporting the discernment of vegetational and environmental drivers for bulk density, carbon and carbon-to-nitrogen ratio. Uncertainty with prediction was also ascertained. A normalized entropy metric constituted from the top two contending groupings was conceived, which better encapsulated prediction uncertainty when compared to conventional entropy. Quantile mapping was also incorporated to uncover insights regarding prediction uncertainty from model ensembles. Structured query language (SQL) was harnessed to efficiently derive and generate rasterized covariates from LiDAR point cloud data. This novel computation consisted of a more detailed digital terrain model (DTM), as well as covariates pertaining to vegetational structure with canopy height model (CHM) and a gap fraction. These covariates were successfully exploited as predictors for modeling with both digital soil mapping and tree species classification.