Formation of Seasonal Vegetation Index Time Series for Information Support of Cropland Monitoring

Alexey Sergeevich Stepanov; Elizaveta Fomina; Konstantin Dubrovin

Alexey Sergeevich Stepanov ХФИЦ ДВО РАН
Elizaveta Fomina Computing Center of the Far Eastern Branch of the Russian Academy of Sciences
Konstantin Dubrovin Computing Center of the Far Eastern Branch of the Russian Academy of Sciences

Keywords: Remote sensing, NDVI, DpRVI, UAV, function fitting, time series

Abstract

Seasonal time series of vegetation indices derived from remote sensing data play an important role in precision agriculture: they are used to study the growth dynamics and condition of crops, forecast yields, and identify crops. Time series formed from optical and radar satellite data, as well as UAV data, often require additional processing to eliminate gaps and enable the calculation of daily values. This article describes the principles of forming seasonal VI series using nonlinear approximating functions. Sentinel-2 and Landsat-8/9 data from 2022 to 2024, Sentinel-1 data from 2021, and monthly DJI Mavic 3M imagery from 2024 were used to assess the fitting accuracy of NDVI and DpRVI series using the following functions: linear combination of Gaussian functions (DG); linear combination of sines (DS); Fourier series (DF); linear combination of logistic functions (DL). Three classes of cropland in the Khabarovsk Krai were considered: soybean, grain crops, and fallow land. It was revealed that for NDVI curves obtained from Sentinel-2 and Landsat-8/9 data, the fitting accuracy based on DF is significantly higher than when using other functions. The approximation of NDVI value series obtained from DJI Mavic 3M data and DpRVI from Sentinel-1 data is possible with various functions with equal accuracy. The average MAPE for the three classes based on Sentinel-2 and Landsat-8/9 data was 14.2% and 7.8%, based on Sentinel-1 data – 12.6%, and based on DJI Mavic 3M data – 7.1%. Based on the results of DF approximation, the main parameters of reference curves for the seasonal progression of NDVI and DpRVI were constructed and determined. It was revealed that the DOY_max values for soybean crops significantly differed from the corresponding indicators for grain crops and fallow land according to optical and radar data. Using the proposed approach to create seasonal time series is one of the elements of automated continuous digital monitoring of arable lands.