Satellite Image Time Series Clustering via Time Adaptive Optimal Transport

Abstract

Satellite Image Time Series (SITS) have become more accessible in recent years and SITS analysis has attracted increasing research interest. Given that labeled SITS training samples are time and effort consuming to acquire, clustering or unsupervised analysis methods need to be developed. Similarity measure is critical for clustering, however, currently established methods represented by Dynamic Time Warping (DTW) still exhibit several issues when coping with SITS, such as pathological alignment, sensitivity to spike noise, and limitation on capacity. In this paper, we introduce a new time series similarity measure method named time adaptive optimal transport (TAOT) to the application of SITS clustering. TAOT inherits several promising properties of optimal transport for the comparing of time series. Statistical and visual results on two real SITS datasets with two different settings demonstrate that TAOT can effectively alleviate the issues of DTW and further improve the clustering accuracy. Thus, TAOT can serve as a usable tool to explore the potential of precious SITS data.

Large-scale unsupervised spatio-temporal semantic analysis of vast regions from satellite images sequences

“…In the current work, we propose a novel methodology that constitutes an advancement in several aspects not previously addressed in depth. Four main contributions are worth mentioning: (i) we design a method to train the embedding considering time series instead of static images, therefore the final embedded vectors contain both spatial and temporal information, (ii) we train the embedding from a grid of satellite images covering a large region, demonstrating the benefits of using embeddings to avoid several technical problems that arise in image fusion such as border effect, inconsistencies between sensors, or temporal shifts when using cloud-free images (Goyena et al 2023), (iii) we go far beyond the classic pixel level in the creation of time series (Guyet and Hervé 2016;Zhang et al 2021) since we use multivariate time series (MTS) of embedded vectors that provide a much richer and more meaningful representation of the spatio-temoral dynamics of the land, (iv) the methodology is fully unsupervised as relies on semantic embeddings and clustering of time series to identify areas, potentially large, that share both similar geographical characteristics and temporal evolution.…”

“…These proposals mainly work at the pixel level or rely on segmentation techniques. Thus, Zhang et al (2021) develop a procedure based on dynamic time wrapping (DTW) distance measures between time series of pixels and Lampert et al (2019) carry out a constrained K-means clustering at pixel level that needs for a proportion of labeled data. Alternatively, Khiali et al (2019) characterize the dynamics of specific objects that evolve similarly, using the well-known segmentation algorithm Mean Shift (Comaniciu and Meer 2002) to identify trackable objects within the SITS.…”