Remote soil observations to predict soil organic carbon

Vahid Khosravi, MSc.

Disertační práce

Abstract:

This thesis presents the results of six separate studies conducted under the topic of “Remote soil observations to predict soil organic carbon (SOC)” spanning from 2021 to 2024. As part of the first objective, the European land use/cover area frame statistical survey (LUCAS) soil spectral library was employed for developing deep learning (DL)-based SOC prediction models. Obtaining results proved the suitability of the large-scale SSL in prediction of SOC specially when leveraging the feature extraction and regression capabilities of the DL algorithms. To accomplish the other objectives, three hundred and twenty soil samples were collected from four agricultural sites in the Czech Republic. An airborne hyperspectral CASI/SASI (CS) imaging campaign was also conducted over the study sites. Several multi- and hyperspectral satellite images related to the study location were also downloaded including Landsat 8-Operational Land Imager (L08-OLI), Sentinel-2 (S2) and Environmental Mapping and Analysis Program (EnMAP). Various machine learning algorithms (ML) were employed to develop SOC prediction models using laboratory, airborne and spaceborne spectra and the results were compared with each other. According to the outputs of the second study, EnMAP data showed enough promise, higher than L08-OLI and S2 data, for prediction and mapping of SOC on the local agricultural topsoil surface. The third and fourth studies were dedicated to considering the impact of green and dry vegetation and soil texture parameters as possible influencing factors on the prediction of SOC. Regarding the green and dry vegetation, the capability of index-based and unmixing-based classification approaches, along with their integration, was evaluated and compared on discrimination of the bare soil pixels of the L08-OLI and S2 data. Based on the results, classification of surface cover using the integrated approach led to more accurate extraction of bare soil and higher performance of SOC prediction models, on both types of satellite data. Effect of the texture parameters was also investigated on the performance of local SOC prediction models developed on S2 and CS hyperspectral airborne data as the main predictor variables. One approach to that objective was to lower the texture variance by stratification of the samples. Therefore, soil samples were segregated based on the i) site-based and ii) texture-based stratification strategies. Overall, texture-based stratification could significantly enhance the SOC prediction, when the texture parameters were added to the S2 and CS data as the main predictor variables. As part of the fourth objective, two modeling approaches: i) ML and ii) regression kriging (RK) by the combination of ordinary kriging (OK) and ML algorithms were used to predict and map SOC using the environmental predictor covariates (EPCs) and thirty genetic algorithms (GA)-selected visible, near-infrared and shortwave-infrared (VNIR–SWIR) wavelengths spectra, individually and combined. Thirty rasters were then created using interpolation of the selected spectra and served as the input variables – with and without EPCs – to test and compare the developed models and SOC predictive maps with each other and with those retrieved from the third approach: iii) kriging using OK of the measured and ML-predicted SOC. Using interpolated spectra as the only input variable yielded a relatively acceptable accuracy. In comparison, the interpolated spectra coupled with EPCs yielded enhanced results as well as maps with less uncertainty than those constructed on the individual datasets. Finally, within the sixth study, a review was conducted to provide quantitative insight into recent trends of SOC digital mapping using remote sensing technology. Accordingly, cropland was the most popular type of land use. Regarding the DSM algorithms, random forest (RF) appeared in the largest number of studies. The terrain and spectral variables derived from the digital el

Abstract:

evation model (DEM) and remote sensing images were the highest demanding among all those used as input predictors. In addition, satellite platforms provided the largest portion of the remote sensing data used for the calibration of DSM models. In summary, the proximal and remote VNIR–SWIR spectroscopy could be deemed as a complementary tool for efficient and sustainable SOC observation.

Keywords

Proximal sensing, Soil organic carbon, Remote sensing, Soil texture, Green and dry vegetation, Digital soil mapping

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 10. 12. 2024

Obhajoba závěrečné práce

• Vedoucí: doc. Asa Gholizadeh, Ph.D.
• Oponent: Lukáš Brodský, externi, Jan Skála, externi, Přemysl Štych, externi

Česká zemědělská univerzita v Praze

Fakulta agrobiologie, potravinových a přírodních zdrojů

Doktorský studijní program / obor:
Agricultural Specialization / Exploitation and Protection of Natural Resources