Stochastic and spectral analysis of time series for identifying periodic land use

Analyse stochastique et spectrale de séries temporelles pour l'identification des occupations du sol

Jean-François Mari (1) , Odile Horn (2) , Marc Benoît (3)

1 ORPAILLEUR - Knowledge representation, reasonning
2 LCOMS - Laboratoire de Conception, Optimisation et Modélisation des Systèmes
3 ASTER Mirecourt - Agro-Systèmes Territoires Ressources Mirecourt

This application note shows the interest of the time series auto / cross covariance feature to mine periodic land use / land cover (LU/LC) in an agricultural landscape. A Bayesian point of view has been adopted to cope with the variability of the real-world data sets. The agricultural landscape is represented by a mosaic of plots, each of them holding a time series of annual LU/LC surveys. An agricultural district is seen as a sample drawn from a time series random process. For each LU, the autocovariance function of the random process is first calculated and its Fast Fourier Transform (FFT) performed. The peaks of the power spectrum determine the frequencies-or the corresponding periods-of this LU. In addition, the analysis of the crosscovariance between two LU/LC may exhibit the co-occurrences of these LU/LC at different lags and determine a return time. In the data mining of agricultural landscapes represented by a mosaic of agricultural plots, results on time series coming from annual LU/LC surveys are presented. They show that the auto / cross covariance coefficients and their spectral representation give immediate and accurate information to a data mining analyst on complex land use rotations and their trends for understanding the underlying logical process in landscape dynamics.

Nous proposons une approche stochastique pour découvir des items périodiques dans un processus stationnaire. Le passage d’une représentation sous forme d’une série temporelle d’items à une représentation sous forme d’une série temporelle de tenseurs multi-dimensionnels nous permet d’utiliser les tech- niques de traitement de signaux multi-dimensionnels. A l’aide des coefficients d’auto corrélation croisés, nous montrons sur des données artificielles qu’une analyse spectrale permet de faire apparaître des comportements périodiques.

Mots clés: Periodic crop; land use succession; Multivariate random process; Time series analysis; Autocovariance

Domaines: Intelligence artificielle [cs.AI]

Jean-François Mari, Odile Horn, Marc Benoît. Stochastic and spectral analysis of time series for identifying periodic land use. 2020. ⟨hal-02944574⟩