Rafael Wiemker

An Iterative Spectral-Spatial Bayesian Labeling Approach for Unsupervised Robust Change Detection on Remotely Sensed Multispectral Imagery

In multispectral remote sensing, change detection is a central task for all kinds of monitoring purposes. We suggest a novel approach where the problem is formulated as a Bayesian labeling problem. Considering two registered images of the same scene but different recording time, a Bayesian probability for ‘Change’ and ‘NoChange’ is determined for each pixel from spectral as well as spatial features. All necessary parameters are estimated from the image data itself during an iterative clustering process which updates the current probabilities.

Investigation on the Torrance-Sparrow specular BRDF model

To describe the directional reflectance characteristics of a specularly reflecting rough surface, Torrance and Sparrow (1967) developed a bidirectional reflectance distribution function (BRDF) model based on geometrical optics. In this study their model is confirmed by a ray-tracing-like simulation, using the same assumptions as in the analytical derivation. The authors also present a simple empirical function that captures the basic features of the analytical model quite well and has been used successfully in BRDF inversion problems including surfaces like roof cover materials and grass canopy.

Modeling the directional reflectance (BRDF) of a corrugated roof and experimental verification

Remotely sensed images with a pixel size of about 1 m can nowadays be acquired by airborne scanners and in the near future also by high resolution satellites. With such a high spatial resolution, remotely sensed data of urban areas can resolve structures like a roof into the different surface segments with different inclinations, e.g. in the case of a gabled roof. The authors have measured the BRDF (bidirectional reflectance distribution function) effects thoroughly on a roof covered with corrugated (sinusoidally shaped) roof tiles and on a sample of flat roof tiles. They modeled the shape of the corrugated tiles by a cosine function and assumed that every infinitesimal surface patch of the roof tile has a BRDF proportional to the BRDF of the flat roof tile. Model results and measurements agree well. The most critical parameters are the ratio height over wavelength of the sinusoidal roof tiles and the intensity of the specular peak of the surface patch. It is possible to retrieve these parameters from the measurements.

Unsupervised Fuzzy Classification of Multispectral Imagery Using Spatial-Spectral Features

Pixel-wise spectral classification is a widely used technique to produce thematic maps from remotely sensed multispectral imagery. It is commonly based on purely spectral features. In our approach we additionally consider additional spatial features in the form of local context information. After all, spatial context is the defining property of an image. Markov random field modeling provides the assumption that the probability of a certain pixel to belong to a certain class depends on the pixel’s local neighborhood. We enhance the ICM algorithm of Besag (1986) to account for the fuzzy class membership in the fuzzy clustering algorithm of Bezdek (1973). The algorithm presented here was tested on simulated and real remotely sensed multispectral imagery. We demonstrate the improvement of the clustering as achieved by the additional spatial fuzzy neighborhood features.

Change detection with 1 m resolution satellite and aerial images

We propose an optimization of a computer based change detection technique based on Iterative Principal Component Analysis (IPCA). We determine and evaluate the changes between an airborne and a spaceborne multispectral image data set, the latter recorded by the commercial satellite IKONOS-2. The change detection algorithm proved to be applicable to large remotely sensed data sets. A vegetation filter, a shadow filter and an oversaturation filter improved the accuracy of the results. When applying all filters more than 80 percent of the objects with changes due to construction activity are detected by the IPCA algorithm. The false alarm rate (change of an object indicated but not verified) is about 5 percent.

Modeling covariance matrices in multitemporal temperature feature spaces

Multitemporal thermal imagery in conjunction with a diurnal temperature model provides a well-known means to derive physical properties of the Earths surface. Classification applications based on such data sets can be used for clustering image pixels with similar heating behaviour. For that purpose, the covariance matrix which describes the mutual dependence of the used features (i.e. temperature values measured at different times of a day) provides an important source of information. In this study we investigate the effect of Gaussian distributed surface properties (thermal inertia, humidity and albedo) on covariance matrices in multitemporal temperature feature spaces with the aid of a diurnal temperature model. The results are compared with experimental findings.

Improved color constant classification of remotely sensed multispectral imagery

For improved multispectral classification and retrieval of Lambertian reflectances from patches of arbitrary surface orientation, the authors investigate the consequences of a dichromatic illumination model accounting for direct sunlight and diffuse skylight. This illumination model leads to the concept of spectral classes as two dimensional planes in feature space. This paper addresses three questions arising from this concept and applies them to experimental data. The author presents the projected spectral angle as a novel spectral distance for the classification of multispectral images. The author shows how the normalized Lambertian reflectance of a surface can be retrieved from at least two observed spectra under arbitrary angles.

The Color Constancy Problem: An Illumination Invariant Mapping Approach

We suggest a novel approach to the Color Constancy Problem for multispectral imagery. Our approach is based on a dichromatic illumination model and filters out all spectral information which possibly stems from the illumination rather than from the reflectance of a given surface. Instead of recovering the reflectance signal, the suggested mapping produces a new only surface reflectance-dependent descriptor which is invariant against varying illumination. Sole input is the relative direct to diffuse illumination spectrum, no assumptions about the possible reflectance spectra are made.