The aim of this thesis is to investigate the potential of computational intelligence techniques for some applications in the analysis of remotely sensed multi-spectral images. In particular, two problems are addressed. The first one is the classification of oil spills at sea, while the second one is the estimation of sea bottom depth. In both cases, the exploitation of optical satellite data allows to develop operational tools for easily accessing and monitoring large marine areas, in an efficient and cost effective way.
Regarding the oil spill problem, today public opinion is certainly aware of the huge impact that oil tanker accidents and oil rig leaks have on marine and coastal environment. However, it is less known that most of the oil released in our seas cannot be ascribed to accidental spills, but rather to illegal ballast waters discharge, and to pollutant dumping at sea, during routine operations of oil tankers. For this reason, any effort for improving oil spill detection systems is of great importance. So far, Synthetic Aperture Radar (SAR) data have been preferred to multi-spectral data for oil spill detection applications, because of their all-weather and all-day capabilities, while optical images necessitate of clear sky conditions and day-light. On the other hand, many features make an optical approach desirable, such as lower cost and higher revisit time. Moreover, unlike SAR data, optical data are not affected by sea state, and are able to reduce false alarm rate, since they do not suffer from the main false alarm source in SAR data, that is represented by the presence of calm sea regions. In this thesis the problem of oil spill classification is tackled by applying different machine learning techniques to a significant dataset of regions of interest, collected in multi-spectral satellite images, acquired by MODIS sensor. These regions are then classified in one of two possible classes, that are oil spills and look-alikes, where look-alikes include any phenomena other than oil spills (e.g. algal blooms...). Results show that efficient and reliable oil spill classification systems based on optical data are feasible, and could offer a valuable support to the existing satellite-based monitoring systems.
The estimation of sea bottom depth from high resolution multi-spectral satellite images is the second major topic of this thesis. The motivations for dealing with this problem arise from the necessity of limiting expensive and time consuming measurement campaigns. Since satellite data allow to quickly analyse large areas, a solution for this issue is to employ intelligent techniques, which, by exploiting a small set of depth measurements, are able to extend bathymetry estimate to a much larger area, covered by a multi-spectral satellite image. Such techniques, once that the training phase has been completed, allow to achieve very accurate results, and, thanks to their generalization capabilities, provide reliable bathymetric maps which cover wide areas. A crucial element is represented by the training dataset, which is built by coupling a number of depth measurements, located in a limited part of the image, with corresponding radiances, acquired by the satellite sensor. A successful estimate essentially depends on how the training dataset resembles the rest of the scene. On the other hand, the result is not affected by model uncertainties and systematic errors, as results from model-based analytic approaches are. In this thesis a neuro-fuzzy technique is applied to two case studies, more precisely, two high resolution multi-spectral images related to the same area, but acquired in different years and in different meteorological conditions. Different situations of in-situ depths availability are considered in the study, and the effect of limited in-situ data availability on performance is evaluated. The effect of both meteorological conditions and training set size reduction on the overall performance is also taken into account. Results outperform previous studies on bathymetry estimation techniques, and allow to give indications on the optimal paths which can be adopted when planning data collection at sea.