Abstract

Thanks to the revisit property of the Earth observation satellites, a huge amount of multitemporal (MT) images are now available in archives. Such kind of images allows us to monitor land surface changes in wide geographical areas according to both long term (e.g., yearly) and short term (e.g., daily) observations. Evolution on the acquisition sensor technology has resulted in the availability of MT and multispectral satellite images with: i) Very High spatial Resolution (VHR) (e.g., Quick-Bird, WorlView-2) and, ii) very high temporal and high spatial resolutions (e.g., Sentinel-2 (S2)). Images acquired by such sensors allow for a detailed geometrical and temporal analysis when compared to medium or high spatial and temporal resolution data. Nevertheless, factors like satel-lite revisit period, the possible competing orders of different users on the satellite pointing (for VHR images only), the limited life of a satellite mission, and weather conditions can lead to: i) lack of enough images acquired by a single sensor to perform MT analysis (VHR case), and ii) lack of regular and continuous Time Series (TS) to perform short term MT analysis (very high temporal and high spatial resolution sensors case) at the level of small objects. Both problems arise from the application requirements on temporal resolution and are being of particular interest in the last years. Two main solutions to the above mentioned problems can be considered: i) use of multisen-sor VHR optical images to replace the missing acquisitions when a single VHR sensor is considered and; ii) development of regression techniques to reconstruct regular and continuous TS from both high temporal and very high temporal resolution sensors. In the literature, most of the MT analysis techniques have been designed to work with: i) VHR images acquired by single sensors, and ii) multispectral images acquired by high spatial resolution sensors, but with low temporal resolution or very high temporal resolution, but with low spatial resolution. Therefore, the effectiveness of ex-isting techniques when applied to the complex MT problems in both VHR multisensor and very high temporal resolution images is reduced. Accordingly, the goal of this thesis is to develop novel tech-niques for the automatic analysis of MT multispectral satellite images such that images acquired by multisensor VHR and very high temporal resolution sensors can be analyzed. The thesis provides four main novel contributions to the state-of-the-art. The first three contribu-tions address the problems arising from the analysis of multisensor VHR multispectral images, whereas the fourth one deals with the problems faced while working with long TS acquired by sen-sors with high spatial and very high temporal resolutions. The first contribution presents an ap-proach for unsupervised CD in multisensor MT VHR images, where the possible sources of noise/changes are studied in detail and a strategy to mitigate them at the levels of pre-processing and feature extraction is presented. In the second contribution, further attention is paid to the ho-mogenization step and a method to generate homogeneous VHR TS focused on the homogenization of intrinsic spectral induced differences is presented. The third contribution further focuses on the detection of multiple changes, while relaxing the knowledge on the statistical distribution of the classes. To this aim, a method based on iterative clustering and adaptive thresholding is imple-mented. Comprehensive qualitative and quantitative experimental results, with real VHR multisen-sor datasets, confirm the effectiveness of the proposed approaches and led to the development of 3 contributions that allow to perform unsupervised bi-temporal CD by means of MT multisensor VHR images. In the fourth contribution, an approach to handle images acquired by both high spatial and very high temporal resolution sensors is presented. To this aim, spectral, spatial and temporal infor-mation of S2 satellite images TS is exploited in different phases and in a fully automatic way, allow-ing for the derivation of different relevant products in the precision agriculture field. Comprehen-sive qualitative and, to some extent, quantitative experimental results confirm the capacity of the method to automatically exploit TS containing both high spatial and very high temporal resolution information. The proposed method can be easily extrapolated for other applications and other sen-sors with similar characteristics to those of S2.