Author:
Stepanov Alexey,Fomina Elizaveta,Illarionova Lyubov,Dubrovin Konstantin,Fedoseev Denis
Abstract
Approximation of the series of the seasonal vegetation index time series is the basis for monitoring agricultural crops, their identification and cropland classification. For cropland of the Khabarovsk Territory in the period from May to October 2021, NDVI and EVI time series were constructed using Sentinel-2A (20 m) multispectral images using a cloud mask. Five functions were used to approximate time series: Gaussian function; double Gaussian; double sine wave; Fourier series; double logistic. Characteristics of extremums for approximated time series for different types of arable land were built and calculated: buckwheat, perennial grasses, soybeans, fallow and ley. It was shown that each type requires a characteristic species. It was found (p<0.05) that Fourier approximation showed the highest accuracy for NDVI and EVI series (average error, respectively, 8.5% and 16.0%). Approximation of the NDVI series using a double sine, double Gaussian and double logistic function resulted in an error increase of 8.9-10.6%. Approximation of EVI series based on double Gaussian and double sine wave causes an increase in average errors up to 18.3-18.5%. The conducted a posteriori analysis using the Tukey criterion showed that for soybean, fallow and ley lands, it is better to use the Fourier series, double Gaussian or double sine wave to approximate vegetation indices, for buckwheat it is advisable to use the Fourier series or double Gaussian. In general, the average approximation error of the NDVI seasonal time series is 1.5-4 times less than the approximation error of the EVI series.
Subject
Artificial Intelligence,Applied Mathematics,Computational Theory and Mathematics,Computational Mathematics,Computer Networks and Communications,Information Systems