Dimensionality Reduction for Remote Sensing Data Analysis: A Systematic Review of Methods and Applications

TL;DR

This paper addresses the challenge of extracting actionable information from increasingly high-dimensional remote sensing data by providing a comprehensive taxonomy of dimensionality reduction (DR) methods and a cross-modality survey of their RS applications. It categorizes DR techniques by dataset, mapping, and preserved properties, and maps them onto RS tasks spanning pre-processing, analysis, and fusion across multiple sensor modalities. The work highlights under-explored DR methods, discusses evaluation metrics, and outlines perspectives such as foundation-model integration and causality-aware DR to guide future RS research. Overall, it offers a practical handbook for selecting DR approaches tailored to specific RS tasks and modalities, with implications for data compression, denoising, visualization, anomaly detection, and predictive analytics.

Abstract

Earth observation involves collecting, analyzing, and processing an ever-growing mass of data. Automatically harvesting information is crucial for addressing significant societal, economic, and environmental challenges, ranging from environmental monitoring to urban planning and disaster management. However, the high dimensionality of these data poses challenges in terms of sparsity, inefficiency, and the curse of dimensionality, which limits the effectiveness of machine learning models. Dimensionality reduction (DR) techniques, specifically feature extraction, address these challenges by preserving essential data properties while reducing complexity and enhancing tasks such as data compression, cleaning, fusion, visualization, anomaly detection, and prediction. This review provides a handbook for leveraging DR across the RS data value chain and identifies opportunities for under-explored DR algorithms and their application in future research.

Figures (16)

• Figure 1: A timeline of common DR methods for RS. DR for feature extraction began with linear multivariate analysis methods like Principal Component Analysis (PCA) hotelling1933analysis in the $1930$s, and became popular in the RS community in the $1970$s lyzenga1978passive. The manifold learning boom began in the late $1900$s, including nonlinear DR like kernel PCA scholkopf1997kernel, and was quickly adopted by the remote sensing community within $10$ years camps2009kernel. Increasing computing power has enabled the popularity of deep learning DR methods, like the variational autoencoder kingma2013auto. Deep learning has been quickly adopted by the remote sensing community zhu2017deep. See Tab \ref{['tab:dr_glossary_standard']} for a glossary of DR abbreviations.
• Figure 2: DR improves performance across the RS data value chain. Data from RS sensors come in various data types, often high-dimensional in spatial, spectral, and temporal dimensions. DR reduces these dimensions to address challenges at each phase of the value chain of RS applications, from pre-processing (e.g., data compression Sec. \ref{['sec:compression']}, cleaning Sec. \ref{['sec:cleaning']}, and fusion Sec. \ref{['sec:fusion']}) to analysis (e.g., visualization Sec. \ref{['sec:visualization']}, anomaly detection Sec. \ref{['sec:anomalies']}, and classification and forecasting Sec. \ref{['sec:predictions']}). The data cube was adapted from teke2013short.
• Figure 3: A taxonomy of DR. DR characteristics are separated into the mapping, dataset, and constraints/property preservation. The DR mapping can be either explicit or implicit. Explicit DR mappings are either linear or nonlinear parameterized functions. Different DR methods are used for different tasks based on the input dataset. Unsupervised DR methods just input a dataset $\mathcal{D} \subset \mathcal{X}$, whereas supervised methods take a dataset of pairs $\mathcal{D} \subset \mathcal{X} \times \mathcal{Y}$ as inputs. The constraints and property preservation for DR algorithms include constraints on the DR mapping, such as matrix factorization or autoencoders, along with properties preserved, including data reconstructions, directions of variance, distributions, geometry, and topology.
• Figure 4: The utility of DR for RS data from various widely used Earth observation sensors. The dimensions of these data are spectral, spatial, and temporal. A higher value for a data type indicates a higher need for DR to reduce features in that dimension. Hyperspectral Imaging (HSI) sensors exhibit high spectral redundancy, requiring dimensionality reduction in the spectral domain. Synthetic Aperture Radar (SAR) and Passive Microwave Radiometer (PMR) data is affected by polarization redundancy and speckle noise, necessitating spatial and temporal DR. Multispectral Imaging (MSI) can have some spectral redundancy. Still, typically, the dimensions to reduce are space and time. Very High Resolution (VHR) sensors often have few bands but a very high spatial resolution. Imaging Radiometers used for meteorology have very long data archives and high spectral resolution. Imaging Spectrometers are primarily used for atmospheric monitoring and can have very high spectral resolution. Finally, Atmospheric Sounder is a recent innovation that can generate thousands of spectral bands at various atmospheric layers.
• Figure 5: DR for RS data compression. Compression reduces a combination of spatial, spectral, and temporal dimensions for increased data processing and transmission rates. The input data from the ambient space is passed through $\phi$ to reduce at least one data dimension, then the data are processed or transmitted in the reduced space, and finally, the data are reconstructed using $\psi$. The grey boxes for the reconstructed data represent the additive reconstruction error to emphasize DR for lossy compression (e.g., $\psi \neq \phi^{-1}$).