Improving the Digital Mapping of Soil Organic Carbon using Environmental Covariates and Machine Learning Algorithms

Abstract

<p>Soil organic carbon (SOC) has been recognized as an important component of functional, healthy and living soils. SOC has been identified to play crucial roles in various ecosystem services, agricultural productivity and soil conservation. Furthermore, the SOC is an essential component of the entire global carbon cycle, occupying the largest share of organic carbon in the entire terrestrial ecosystems. The position of SOC in the global carbon cycle and its potential roles for influencing the climate change scenarios by acting either as a source or a sink of atmospheric carbon dioxide have brought the soils into the limelight globally in recent decades. The depletion of SOC, particularly in agricultural lands, is a major environmental issue in many parts of the world. It is important to understand the spatial distribution of SOC in order to address such environmental concerns. Therefore, there is increasing interest around the world in mapping the distribution of SOC. The development of various allied fields such as quantitative approaches to map soil attributes, availability of a large volume of better resolution covariates and the accessibility of high-performance computing facilities have enhanced the growth of soil mapping technologies over the last two decades in the framework of digital soil mapping (DSM). This research aimed at improving the predictive mapping of topsoil SOC contents, especially in the croplands of Nepal. Due to the difficult terrain conditions in most of the geographic areas of Nepal, intensive and frequent soil surveying is not practical. We have used available legacy soil information, recent soil data and other ancillary datasets from various sources to achieve this aim.</p> <p>Our review work, based on the systematic evidence mapping approach, unveiled the current state of the progress in mapping the spatial distribution of SOC using DSM. There was an uneven geographical distribution of the DSM studies across the world that focused on the mapping of SOC contents and stocks. The studies were clustered in some countries such as China, Australia and the USA, with the highest number of articles being published in 2016-2017 over the period of study (2013 to 2019). Not a single publication was found in Nepal that used the modern DSM approach. So the further study was focused on improving the predictive mapping of SOC contents in Nepal. The review also unearthed the trend of predictive models that were progressing from linear models to the machine learning (ML). Most of the comparative studies had claimed the Random Forest (RF) model to be a better predictor of SOC than other models. The predictor models combined with the interpolation of the prediction residuals in the framework of regression-kriging were reported to be even better than the individual models alone by some studies. However, no single model was reported to be the best one in all circumstances, heralding a need to assess the performance of promising algorithms using sufficient field and lab based SOC observation data. The correlation between environmental variables and soil carbon levels was observed to be mostly dependent on environmental circumstances, soil depth, mapping resolution, and the size of the area under consideration. Climate was reported to be the most critical component in SOC levels for mapping at regional scales, followed by parent materials, terrain, and land use. However, heterogeneity in land use was reported to be more significant in predicting SOC for mapping at a scale that represented plots or small fields. Our study reveals that the factors representing the ‘organisms' component are among the most common among the top five covariates, followed by variables representing the ‘climate' factor and then ‘topography’. In comparison to an earlier study, this study reveals that the practices of validating the prediction results and quantification of spatially explicit uncertainty have become more common in recent DSM studies for mapping SOC. However, the use of additional probability sampling approach for evaluating the prediction is still not common, probably due to the additional costs associated with it. Comparison of the RF and a linear model combined with kriging of residuals in an alluvial plain area of the Terai Physiographic region of Nepal revealed some interesting patterns of topsoil SOC driving factors in this area. In this fertile and intensively cultivated warm subtropical plain area, the mean SOC content was 10.98±0.01 g kg-1. The SOC contents were also predicted and quantified for different soil and land cover types. The RF model, which performed better than the linear regression kriging, revealed that the proximity to the river systems and the deposition of silts were the most prominent driving factors of topsoil SOC. Controlling silt deposition from upstream hills could thus affect the success of initiatives to raise SOC levels. </p> <p>Evaluation of multi-season satellite imagery in a montane ecosystem revealed that the prediction of the topsoil SOC was improved when combined with relevant topographic variables. However, if a parsimonious model is preferred, an image from the right timing in the cropping calendar could also yield comparable results. Therefore, in the national scale SOC mapping, satellite image-based indices from a single season were used for the efficiency of the model to run at a high resolution to cover the entire country. For the current study, the training data were collected through conditioned Latin Hypercube Sampling technique for the efficiency of sampling and an additional probability sampling was carried out to perform genuinely independent validation of the prediction results. </p> <p>In the framework of DSM, the data layers representing soil types and soil attributes can be used as the predictors for other soil attributes. This was also revealed by the systematic evidence mapping or the review work in Chapter 2. As the existing soil type map for Nepal was available with less detailed information with composite mapping units, a modern DSM approach was employed to improve the legacy map to get more detailed soil map. The spatial disaggregation method was found to improve the details and accuracy of the existing legacy soil map. The output of this study was used as a predictor of SOC contents in the next study for the predictive mapping of topsoil SOC contents for the entire country.</p> <p>Four promising machine learning models were built and compared using a large volume of latest field and laboratory-based point SOC observation data from the agricultural lands and selected relevant covariate layers to predict and map SOC contents across the entire country of Nepal. The choice of the models, covariates and techniques in this study was informed by all of the previous studies explained above. The predicted results in this study were also compared against a global SOC contents dataset. The results of all the models built in this study were better than the global dataset. The prediction of the RF model was better than other models, closely followed by the Cubist model. The prediction of the extreme gradient boosting was also competitive. The baseline SOC contents map produced using the best model in this study provides a useful spatial soil information for the croplands of Nepal. </p> <p>These studies have led to the improvement of SOC prediction in Nepal using state-ofthe-art DSM techniques within the constraints of available data. The studies have also provided a reference for improving SOC contents in similar physiographic settings in other parts of the world as well. The disaggregated and improved World Reference Base soil map is also a novel and useful spatial information product for land management decisions across various environmental and agricultural issues for the country. The new topsoil SOC baseline map is expected to help prioritise SOC content enhancement initiatives in the croplands of the entire country. The use and evaluation of various predictive algorithms and covariates is expected to assist the prospective researchers and practitioners in this field to guide a more informed selection of the prediction models and covariates for the predictive mapping of SOC. </p>