## Edge detection

A segmentation approach that is dual to finding the homogeneous interior of a region is to determine where the region boundaries are located. This strategy is usually implemented using edge detection methods, which look for local discontinuities in the grey level distribution of an image. Since in general the strongest edges coincide with region boundaries, edge 1Note that step 2(ii) of Schoenmaker's Algorithm 5.4 is incomplete a correct formulation of the set is CSnbr (Su, Sk) (Su, Sk) G CS A...

## Simulation of remote sensing images

The generation of artificial satellite images can be done in many ways with varying degrees of realism. A method which preserves both first- and second-order statistics has been described by Schouten and Klein Gebbinck 92 and is briefly summarised in this section. The first step in generating an artificial image is to draw the boundaries of objects such as agricultural fields, roads, and ponds on a white piece of paper using a black pen. Figure 6.2(a) shows an example, which we used as the...

## Principal component analysis

Traditionally, the objective of principal component analysis (PCA) is to reduce the dimensionality of a data set while retaining as much of the relevant information as possible. This goal can be achieved by rotating the coordinate system such that most of the variation in the data is found along a limited number of axes, the so-called principal components. The axes where the data shows little or no variation are disregarded, which corresponds to restricting the original feature space to a...

## Spectral confusion

Based on the available supervised data, spectral confusion was the only source of inaccuracy of DDD and FBC that could studied. In this thesis, the term spectral confusion is used to describe the resemblance of one (mixture of) ground cover type(s) to another (mixture of) ground cover type(s). Basically, we distinguish three different kinds of spectral confusion. The first sort occurs during decomposition of mixed pixels based on the linear mixture model if several linear combinations of...

## Setup of the experiment

The study site that was selected is located near the village of Zeewolde in the south of the Flevoland polder. The land, which was reclaimed from the lake named IJsselmeer in 1968, is flat with an altitude of 3.30 m below sea level and is mainly used for the cultivation of arable crops. A network of roads, canals, and ditches divides the area into lots, which are basic units of approximately 500x1600 m2 that can be rented or bought. Many farmers subdivide their lots into several agricultural...

## A little comparative experiment based on MAIS data

From the previous section it has become clear that the concept of ILS is simple and easily extendible to an arbitrary number of classes. Furthermore, we have shown that the method has the potential to be fast as its complexity depends on the number of classes instead of the number of spectral bands. However, it must still be checked that the number of iterations required is not too large for ILS to be applicable in practice. Besides that, the accuracy of the fractions vector estimated by ILS...

## Brute force approximation

To understand the complex relations between the ei, a look at the n-dimensional feature space is helpful. According to convex geometry principles, pixels that are composed of c Figure 3.3 Maximum likelihood solution of the physical linear mixture model. To solve the model, one point pi per endmember has to be determined, whose Mahalanobis distance to mi is minimal. The sum-to-one constraint demands that these points lie on a line through x the positivity constraint requires that pi and p2 lie...

## Results of edge detection

The performance of the Frei-Chen edge detector is dependent on the setting of its threshold. Figure 5.8 shows the specificity, sensitivity, and average percentage correct achieved for a wide range of threshold settings5. As expected, at low thresholds many pixels are marked as edge and therefore mixed pixels, which leads to a high sensitivity but a low specificity. If the threshold is raised, the sensitivity decreases while the specificity increases, ultimately reaching 100 . The best...

## Neural networks

In the last decade neural networks have emerged as a powerful classification tool for, amongst others, remotely sensed image data (Benediktsson et al. 11 , Hepner et al. 47 , Kanellopoulos et al. 57 ). An artificial neural network (ANN) is a system of highly connected but simple processing units called neurons, which is inspired by the architecture of the human brain. Calibration of an ANN is done by repeatedly presenting a set of training examples to it, which has the advantage that the class...

## Introduction to remote sensing

The field of remote sensing is very broad and has been described from many angles by numerous authors, e.g. Campbell 19 , Lillesand and Kiefer 66 , and Sabins 88 . In his book, Campbell tried to identify the central concepts of remote sensing and came up with the following definition Remote sensing is the practice of deriving information about the earth's land and water surfaces using images acquired from an overhead perspective, using electromagnetic radiation in one or more regions of the...