Bathymetry Estimation from Multi-Spectral Satellite Images Using a Neuro-Fuzzy Technique

The aim of this work is the study and the development of a technique for bathymetry estimation, which is based on the exploitation of information contained in optical images, collected by satellite sensors. The use of satellite images allows to inspect a wide geographical area, and to produce the corresponding bathymetric map, in an economical way. Since remote areas can easily be observed by satellite sensors, remote sensing techniques could represent a useful support to Rapid Environmental Access (REA) activities, that consist in getting information on hardly accessible zones. Moreover, the exploitation of high resolution satellite sensors data, such as Quickbird data, allows to describe the coastal zone with high accuracy, although this implies a reduction of the spectral information. In this work we propose an accurate supervised method based on the use of a neuro-fuzzy system, whose input is made of only three spectral bands. The method consists in the application of an Adaptive-Network-based Fuzzy Inference System (ANFIS) to the optical satellite image of the area of interest. We applied the technique to two Quickbird images of the same area, acquired in different years and in different meteorological conditions. In particular, the first one has been acquired in calm sea conditions, and is supplied with a large dataset of in-situ measured depths for the training and the validation of the method. The second image has been acquired in slight sea conditions and is supplied with a limited dataset of in-situ measured depths, collected along two transepts in the scene. These two cases allow to study and compare the performance of the presented technique, taking into account the effect of both meteorological conditions and training set size reduction on the overall performance. On the first image we achieved a mean STD of 36.7 cm for estimated water depths in the range [-18,-1] m. We then studied the performance of the method in realistic situations of limited in-situ data availability, that is using as a training set only data collected along closed paths in the same image. In this case we obtained a mean STD of 45 cm. In addition, we studied the effect of limited data availability together with unfavorable sea conditions by applying the method to the second image. In this latter case we achieved a mean STD of about 64 cm, which is still a good result.