The assessment of the influence of land-use and topography on the spatial distribution of soil organic carbon using digital soil mapping

Shahin Nozari

Doctoral thesis

Abstract:

This Ph.D. thesis investigates the intersection of digital soil mapping (DSM) with geographical information systems (GIS) and spatial statistics. Utilizing digitized data from legacy soil survey maps, the research investigates soil organic carbon (SOC) distribution in the Czech Republic, with a focus on the Liberec and Domažlice districts. Basically, this study aims to understand factors influencing SOC spatial distribution across various land-uses and topographic variables. It contributes to advancing DSM knowledge, refining local maps for Liberec and Domažlice regions, and improving the comprehension of SOC spatial patterns. The research also presents a spatial prediction structure for SOC using DSM and compares machine learning (ML) models. Key objectives of this study included investigating the impact of ML model choice on SOC prediction accuracy, emphasizing the importance of training data quality, and indicating that the suitability of covariates such as Normalized Difference Vegetation Index (NDVI) and land cover is comparable to digital elevation models (DEMs). This study facilitates improvements in ML model selection and training data progress in DSM, fostering enhancements in local maps for areas with similar climates. To achieve these objectives, in this study, the correlation between slope, elevation, and clay on SOC content in different land-uses was explored. Subsequently, the effects of region and elevation on predictors in SOC prediction models were evaluated. Following that, as the primary objective of this research, DSM and ML methods, including Random Forest (RF), Quantile Random Forest (QRF), and Cubist, were utilized to generate SOC maps at a 100 m spatial resolution, enabling the prediction of SOC distribution in the aforementioned districts in the Czech Republic. Additionally, infrared spectroscopy was used to assess the effects of land-use on soil organic matter (SOM) quality. Analyzing the relationship between slope, elevation, and clay with SOC across different land-uses, the study found that the correlation between SOC and clay lacked statistical significance, despite both increasing with slope. Altitude positively correlated with SOC, though the correlation was only moderately strong. Multiple regression models revealed stronger SOC correlations in the Domažlice district, indicating a significant impact of land-use variations on SOC distribution. In summary, the integration of slope, elevation, and multiple regression analysis enhanced SOC predictions. Examining the impacts of region and elevation on SOC prediction models revealed that edaphic series and soil classes are robust predictors at lower altitudes, while higher altitudes highlighted the significance of topography-related predictors. The meticulous selection of influential covariates was recognized as pivotal in DSM. Optimal prediction outcomes were noted in smaller, but consistent regions, such as specific natural forest areas, albeit with acknowledged model failures in certain natural forest areas. The study identified model limitations related to uncertainties arising from data harmonization, transformation, and standardization. Finally, this research challenges the common reliance on DEMs by highlighting the equally crucial roles of other covariates, such as the NDVI and land cover. The results also showed that RF model consistently outperforms QRF and Cubist models in both districts, demonstrating better accuracy in predicting SOC. The RF model, leveraging diverse terrain covariates, exhibited greater R2 values, smaller RMSE and MAE values, and a more uniform uncertainty distribution. Key predictors for SOC included land cover (vegetation) and elevation, with forest-dominated areas displaying the highest predicted SOC content. Notably, coniferous forests showed the greatest uncertainty in SOC predictions, likely due to a limited number of samples and increased covariate variability. The results obtained from infrared spectroscopy hi

Abstract:

ghlighted the significant influence of land-use on SOM composition. Forest soils showed more pronounced variations, with the upper layer showing intensified aliphatic bands (3010–2800/cm) and higher acidity in SOM, evidenced by carboxylic band intensity, compared to grassland and cropland. Notably, grassland had distinct fulvic acids (FAs) compared to other land-uses, and in cropland soils, the aromaticity of humic acids (HAs) increased with depth. These findings underscore the crucial role of land-use in shaping SOM composition, emphasizing the necessity of considering diverse land-use factors for accurate predictions of SOC. Recommendations for future work include exploring advanced ML approaches, such as deep learning, for improved SOC predictions, and conducting integrated environmental impact analyses to understand the joint effects of climate, vegetation patterns, and topographical factors on SOC distribution. Additionally, in-depth studies on microbial dynamics in SOC, longitudinal soil health assessments, and community-collaborative research initiatives are suggested to enhance understanding of regional land-use practices and environmental transformations.

Keywords

Digital Soil Mapping, Soil Organic Carbon, Machine-Learning, Statistical Analysis, Random Forest, Quantile Random Forest, Cubist, Geographical Information Systems, Diffuse Reflectance Infrared Spectroscopy with Fourier Transformation (DRIFT)

Language used: English

Date on which the thesis was submitted / produced: 11. 1. 2024

Thesis defence

• Supervisor: prof. Dr. Ing. Luboš Borůvka
• Reader: Vít Penížek, Ph.D., doc. Ing., Radim Vácha, externi, Vítězslav Vlček, externi

Czech University of Life Sciences Prague

Faculty of Agrobiology, Food and Natural Resources

Doctoral programme / field:
Agricultural Specialization / Exploitation and Protection of Natural Resources