Uncertainty and results stability of three digital soil mapping algorithms applied to the soil cover of a farm situated on the north of Udmurt Republic, Russian Federationтезисы доклада
Дата последнего поиска статьи во внешних источниках: 26 декабря 2017 г.
Аннотация:One could generate multiple soil maps for the same territory with advanced DSM algorithms.
Some areas will belong to the same classifcation unit on all maps produced by the algorithm.
Such areas may be named “zones of stability”. The other areas will be classifed differently with
the same algorithm. The aim of this study was to identify which territory of the study area are
zones of stability with a probability of 100%, 95% and 75%. Traditional soil map (under Russian
classifcation) was done for a farm situated on the North of Udmurt Republic of Russia. Eight
soil types and subtypes that are typical for the south taiga zone for the Central Russian Plain
were detected on the territory of 19 square km. Several digital soil maps (100 for an algorithm)
were produced using multinomial logistic regression, support vector machine and random forest
algorithms on the base of covariates: legacy data, terrain attributes and land use. Independent
random sample was used to test all soil maps. Traditional soil map was the most accurate and
gave 82% of right profles classifcations. MNLR, SVM and RF algorithms gave less than 70% of
right profles classifcations. Maps of 100%, 95% and 75% stable soil diagnostics were produced
for each algorithms. Three algorithms showed similar results. The proportion of 100% stable land
areas was about 40%. Stability zones matched the arable lands on the watersheds. Territories
under forests and areas situated in ravines and gullies showed the highest level of uncertainty of
classifcation using digital soil mapping algorithms. Slope territory manifested the intermediate
level of uncertainty