ADAPTIVE APPROACH TO THE IMPUTATION OF ONE-DIMENSIONAL TIME SERIES OF SATELLITE DATA BASED ON MACHINE LEARNING METHODS

Khannanov Nail Kamilevich

Modern satellite observation systems form continuous time series of measured parameters, which are used in the tasks of analyzing and predicting the dynamics of observed phenomena, however, the presence of gaps resulting from equipment failures, interference in communication channels, etc. increases the level of uncertainty in the interpretation of satellite data and reduces the effectiveness of sources of information of this kind. Simple linear interpolation, classical statistical models, and neural network models are most often used in such situations. However, classical interpolation schemes ignore the complex dynamics of signals and have significant methodological errors, as a result of which they reduce accuracy on long fragments of gaps. Statistical models are sensitive to the assumption of stationarity and accumulate error in the long absence of data. Neural network models demonstrate excessive complexity for short passes with high demands on computing resources. The paper considers an adaptive method of imputation of one–dimensional satellite time series, which includes preprocessing and comparing three imputation tools - linear interpolation, an autoregressive integrated moving average model, and a recurrent neural network. It is shown that with short gaps, simple methods provide a quality comparable to a neural network, but with an increase in the length of gaps, the neural network demonstrates a stable advantage. The values of the coefficient of determination remain positive and retain practical value, whereas the autoregressive model and linear interpolation are characterized by an increase in the average absolute error and a decrease in the coefficient of determination up to negative values. Based on the revealed dependencies, an algorithm for selecting a method based on the length of the pass is proposed, which forms the basis for adaptive imputation of satellite time series.

Keywords:Data imputation; time series; satellite data; linear interpolation; statistical models, recurrent neural networks; adaptive method selection.