Manifold Embeddings for Multispectral Time-Series Land Disturbance Detection

Dimensionality reduction techniques have been successfully applied in remote sensing to reduce redundant information. However, achieving dimensionality reduction and lossless recovery for multispectral data at any global location remains a challenge, particularly given the complex and variable nature of surface conditions. Furthermore, it is still unclear if the reduced features maintain temporal continuity and can be effectively integrated with existing time series algorithms for disturbance detection. This study trains a Uniform Manifold Approximation and Projection (UMAP) model based on Harmonized Landsat Sentinel-2 (HLS) imagery to accomplish multispectral dimensionality reduction. Subsequently, the manifold embeddings are used in the Continuous Change Detection and Classification (CCDC) algorithm for land disturbance detection. Two key conclusions are drawn from this study: 1) a general multispectral dimensionality reduction model was constructed based on UMAP, which is applicable to all global land surfaces and any seasons. The manifold embeddings exhibit a stable value range and preserve the coherence of the time series. 2) compared to full-spectrum multispectral data, the manifold embeddings achieved comparable performance in image prediction and disturbance detection. Our study demonstrates the potential of manifold learning-based representation of global land surface reflectance spectra for lightweight storage and processing of dense satellite image time series, while keeping the ability to detect any kinds of land disturbance.