|
Keywords
|
remote sensing, machine learning, aggregates fraction, spatial distribution, sustainable soil management
|
|
Abstract
|
Spatial distribution of soil organic carbon (SOC), and
total nitrogen (TN) contents in different aggregate fractions
(DAFs) were investigated by applying multiple
machine learning models (MLMs) i.e., cubist (CB),
support vector regression (SVR), and random forest
(RF), along with environmental variables in the framework
of digital soil mapping (DSM). One hundred
samples were taken from the soil surface layer (0−15
cm) in the Aji-Chai watershed, in northwestern Iran. TN
and SOC were measured in three soil aggregate sizes
(macro, meso, and micro-aggregates). Among the studied
machine learning models (MLMs), the RF model revealed
exceptional performance and the lowest uncertainty for predicting SOC and TN contents in DAFs. The R2 values for the prediction of SOC in
DAFs were 0.86 for SOCmacro, 0.83 for SOCmeso, and 0.81 for SOCmicro. For the TN content in different fractions, the R2 values were ordered as
0.70 for TNmacro, 0.71 for TNmeso, and 0.73 for TNmicro, respectively. Variable importance analysis (VIA) results indicated that factors like vegetation
indices such as Corrected transformed vegetation index (CTVI), and normalized difference vegetation index (NDVI), followed by topographic
attributes, had a substantial impact in exploring SOC and TN contents in DAFs. In macro-aggregates, the highest SOC and TN contents were
found in dense pasture, semi-dense vegetation, and orchards. Conversely, in meso- and micro-aggregates, the lowest contents were observed
in rainfed agricultural lands, sparse pastures, and barren regions, respectively. The modeling results open new windows in the field of soil fertility
and physics intending to link the content of SOC and TN variation in DAFs. Ultimatly, the RF model demonstrates strong predictive capabilities
for SOC and TN contents in DAFs, achieving impressive R² values. Influential factors include vegetation indices and topography. The resulting
prediction maps significantly enhance spatial planning and guide sustainable land management practices, effectively linking soil quality indicators
to specific land-use types for improved soil health.
|