Machine Learning application to predict and map Soil Organic Carbon in the Bohemian Highlands (Czech Republic)

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00027049%3A_____%2F24%3AN0000103" target="_blank" >RIV/00027049:_____/24:N0000103 - isvavai.cz</a>

Výsledek na webu

—

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Machine Learning application to predict and map Soil Organic Carbon in the Bohemian Highlands (Czech Republic)

Popis výsledku v původním jazyce

We applied the Digital Soil Mapping (DSM) approach to map soil organic carbon (SOC) content across Krasna Hora Nad Vltavou, situated in the Příbram District, Central Bohemian Region, Czech Republic. We focused on two soil layers: 0-10 cm (SOC 10) and 0-30 cm (SOC 30), employing various machine learning models. The study area covers approximately 500 km², characterized by elevations ranging from 268 to 574 m a.s.l., with geomorphologic classification as a Bohemian highland featuring gently undulating mountains. Predominant land cover primarily consists of arable lands with some forested areas. Utilizing data from 105 georeferenced soil profiles sampled by horizons, we analysed soil samples in the laboratory for organic carbon and total nitrogen using dry combustion. Four machine learning models Random Forest (RF), Multivariate Adaptive Regression Splines (MARS), Support Vector Regression (SVR), and Elastic Net (ENET) were employed for modeling and prediction, using environmental covariates such as geomorphometric parameters, climatic variables, and mosaic bare soil indices. Model performance assessment through 10-fold cross-validation revealed that RF consistently outperformed other models, exhibiting the lowest Mean Absolute Error (MAE) values and the highest R-squared (R²) scores for both SOC 10 (R² 0.79, MAE 0.32%) and SOC 30 (R² 0.82, MAE 0.28%). While MARS provided acceptable results, ENET and SVR showed higher Root Mean Squared Error (RMSE) and lower R² values, indicating lower accuracy in SOC modeling and prediction. The resulting SOC 10 map displayed a range from 1.73% to 7.76% with a mean SOC content of 2.72%, while SOC 30 exhibited a narrower range (1.22% to 3.13%) with a mean value of 1.73%. This research underscores the effectiveness of machine learning in DSM for predicting and mapping SOC across diverse land cover types, emphasizing the influence of temperature and precipitation on the spatial distribution of SOC content. The findings contribute valuable insights into understanding the potential impact of climate change on the future distribution of SOC in Bohemian uplands.

Název v anglickém jazyce

Machine Learning application to predict and map Soil Organic Carbon in the Bohemian Highlands (Czech Republic)

Popis výsledku anglicky

We applied the Digital Soil Mapping (DSM) approach to map soil organic carbon (SOC) content across Krasna Hora Nad Vltavou, situated in the Příbram District, Central Bohemian Region, Czech Republic. We focused on two soil layers: 0-10 cm (SOC 10) and 0-30 cm (SOC 30), employing various machine learning models. The study area covers approximately 500 km², characterized by elevations ranging from 268 to 574 m a.s.l., with geomorphologic classification as a Bohemian highland featuring gently undulating mountains. Predominant land cover primarily consists of arable lands with some forested areas. Utilizing data from 105 georeferenced soil profiles sampled by horizons, we analysed soil samples in the laboratory for organic carbon and total nitrogen using dry combustion. Four machine learning models Random Forest (RF), Multivariate Adaptive Regression Splines (MARS), Support Vector Regression (SVR), and Elastic Net (ENET) were employed for modeling and prediction, using environmental covariates such as geomorphometric parameters, climatic variables, and mosaic bare soil indices. Model performance assessment through 10-fold cross-validation revealed that RF consistently outperformed other models, exhibiting the lowest Mean Absolute Error (MAE) values and the highest R-squared (R²) scores for both SOC 10 (R² 0.79, MAE 0.32%) and SOC 30 (R² 0.82, MAE 0.28%). While MARS provided acceptable results, ENET and SVR showed higher Root Mean Squared Error (RMSE) and lower R² values, indicating lower accuracy in SOC modeling and prediction. The resulting SOC 10 map displayed a range from 1.73% to 7.76% with a mean SOC content of 2.72%, while SOC 30 exhibited a narrower range (1.22% to 3.13%) with a mean value of 1.73%. This research underscores the effectiveness of machine learning in DSM for predicting and mapping SOC across diverse land cover types, emphasizing the influence of temperature and precipitation on the spatial distribution of SOC content. The findings contribute valuable insights into understanding the potential impact of climate change on the future distribution of SOC in Bohemian uplands.

Druh

O - Ostatní výsledky

CEP obor

—

OECD FORD obor

40104 - Soil science

Projekt

—

Návaznosti

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů