Improving accuracy in snow depth modeling based on the effectiveness of selecting index features in parametric and non-parametric regressors

Achieving the spatial distribution of snow depth should be done through observation and on a compact scale. But due to practical limitations, it is difficult and sometimes impossible to collect information, especially in the mentioned scales. According to the existing problems, the use of machine learning approach and feature selection can increase the applicability of snow depth zoning in high areas. In this research, the effectiveness of reducing ineffective features in learning based on parametric and non-parametric models has been investigated. The samples used to test the hypotheses were taken from Chalgerd region of Iran. For this purpose, first, using the hypercube method, the location of 100 specific points and during a field operation, snow depth data were collected at the desired points and also at 195 other points randomly and with the federal model sampler. Then, using digital height model, 25 geomorphometric parameters were extracted and together with 6 bands of Landsat 8 satellite images and NDSI index were selected as model inputs. The results of this research showed that the parametric and non-parametric methods did not have acceptable accuracy in modeling the snow depth, but the linear regression model with the forward greedy feature selection method and the particle population optimizer with the mean squared error equal to 22.17 and 22.19 were able to Model snow depth changes with better accuracy.