Page 105 - Çevre Şehir ve İklim Dergisi İngilizce - Özel Sayı

Mustafa Sert - Abdullah Emin Akay
Ayhan Ateşoğlu

Medium Term Weather Forecasts (ECMWF) was utilized as meteorological
data (Hersbach et al., 2020). Vegetation cover and land use data were produced
from the CORINE Land Cover dataset via the Copernicus Land Monitoring
Service. This set was used in the artificial intelligence (AI) stream to filter out
the Land Use Land Cover (LULC) classes that do not exactly correspond to
agricultural areas.
In order to be used in ground-based measurements, 84 soil samples
were taken from agricultural lands and the Soil Organic Carbon (SOC)
content was derived by the Walkley-Black method (Walkley and Black, 1934).
Additionally, soil textures were determined by the Bouyoucos hydrometer
method (Bouyoucos, 1962). An AI-based approach was adopted to produce
soil organic carbon and soil texture maps from a digital soil map (Wadoux
et al., 2020). The Sentinel-2 satellite images, terrain data from AB- DEM
and climate data from ERA5 were used as input to the model. The XGBoost
(eXtreme Gradient Boosting) algorithm using decision trees (DT) was utilized
to optimise the model (T. Chen and Guestrin, 2016).
A soil erosion layer was derived based on an artificial intelligence
architecture, open source EO data and an improved soil layer generated by
the RUSLE method. The RUSLE method was adopted to estimate the annual
soil loss. According to the method, the interaction of R, K, C and LS factors
reveals the amount of soil lost as a result of erosion (Renard et al., 1997). The
accuracy of the models was achieved by AI model performance measurements
which include the following:
• The R2 coefficient (coefficient of determination) was applied to quantify
the degree of any linear correlation between the output revealed and
the one estimated by the model.
• The root-mean-square deviation (RMSE) was used to determine the
average difference between the estimated values and the actual ones.
The study demonstrated the importance of spatial resolution and the
importance of integrating artificial intelligence in RUSLE to find the erosion
status and to create a soil erosion map. The use of soil organic carbon affecting
the factor K and texture maps is an important novelty of the study. The learning
algorithm XGBoost demonstrated a reasonable accuracy with high resolution
maps providing detailed spatial distribution of soil erosion indicators.

4. Conclusions and Recommendations

Today, in order to take measures against the environmental problems
increasing with climate change, it is of utmost importance to identify their
source and severity. Erosion and desertification are among the leading

92 Journal of Environment, Urban and Climate