Unsupervised machine/deep learning mapping (clustering) for single/multi-source remote sensing data

In recent years, there has been an explosive growth in remotely sensed (RS) data usage in geoscience and Earth observation applications. The recent technical advancements allow users to acquire various types of rich information. A quintessential example of RS data is hyperspectral imagery. A hyperspectral image (HSI) provides rich spectral data over a wide range of the electromagnetic spectrum. Such information enables users to identify, track and distinguish different materials and objects. Another RS data example is light detection and ranging (LiDAR). LiDAR yields information on the altitude of the observed objects. Hence, it allows distinguishing objects that might share common spectral characteristics but have different altitudes (e.g., tree species). Although RS data have high potential for material characterization, the processing of such data poses challenges, for example, on the high dimensional nature of the datasets or an efficient fusion of multiple data sources. Machine learning approaches march as the pioneer solutions to the aforementioned issues. Among machine learning approaches, supervised learning approaches perform accurately in various tasks (e.g., classification and regression). Nevertheless, such approaches demand an immense number of training samples during their process. Specifically, in geoscience and Earth observation applications, acquiring training samples is a labor-intensive, time-consuming, and expensive task. Moreover, in some cases, it is not possible to generate training samples due to limited accessibility. Therefore, supervised approaches constitute a shortcoming with respect to the availability of training samples. On the contrary, unsupervised learning approaches accomplish different tasks (e.g., feature extraction and clustering) by merely analyzing the data itself. To be more specific, the clustering problem refers to grouping similar pixels into clusters. In general, clustering approaches can be split into two categories: (1) Conventional shallow learning (CSL) and (2) Deep learning (DL)-based approaches. This study is devoted to the development of unsupervised CSL and DL approaches for single- and multi-sensor remote sensing data clustering.

Publication year:2023

Keywords:Doctoral thesis

Accessibility:Open