Rule-Based Classification of Water in Landsat MSS Images Using the Variance Filter

The variance filter is a textural algorithm capable of distinguishing between pat, uniform water bodies and cloud or mountain shadows when applied to satellite imagery. The filter output forms the basis of rules used by a knowledgebased classifier, which segments water adaptively. The use of the filter in the unsupervised classification of water is demonstrated on two spectrally varied Landsat MSS images. The same images are segmented using a conventional thresholding algorithm. The two algorithms identify a similar proportion of the water pixels in both images; however, the rules-based algorithm does not generate any false positives, whereas the threshold algorithm misclassifies many shadow pixels as water. The rules-based algorithm is less efficient at finding small lagoons and swamps than a t finding large water bodies. Introduction This paper describes one algorithm which has been developed as part of a long term project. The long term goal of the research is to devise methods of partially analyzing Landsat data without human intervention or guidance. The rationale for this is the need to solve what is sometimes referred to as the terabyte problem. There is a need to process, by computer, the very large volumes of data being received from Earth observation satellites, given that there is a shortage of trained analysts. The short term goal of this work is to reliably identify (segment) pixels of water in a Landsat multispectral scanner (MSS) image, and without obtaining any false positive pixels. Cloud shadow and western facing mountainsides are examples of features which are often confused with water by classifiers. The author has elected to use any algorithm which will provide useful results; the range of aigorGhms investigated in I the project to date includes statistical methods, texture algoI rithms, contextual algorithms, rules, production systems, fuzzy logic, genetic algorithms, and neural networks. This paper is confined to a description of one texture algorithm, two contextual algorithms, and several rules. The rules were i developed by trial and error in order to overcome several difficulties and to meet several criteria which are described in the next section. The experimental work has been carried out on Landsat MSS images because the data sets are less costly than Thematic Mapper (TM) data, and MSS images continue to be widely used for land-use and resources studies. Preliminary tests with TM images suggest that the algorithms will be equally successful with that type of data. Traditionally, segmentation of water is carried out under human supervision using spectral information. In studies of the Great Barrier Reef Marine Park, Australia, Jupp et al. School of Electrical and Electronic Systems Engineering, Queensland University of Technology, P.O. Box 2434, BrisI bane 4001, Australia ([email protected]). Wilson (1985) state, "Using Band 7, each image is separated into water and other areas (such as land and clouds, etc.) by a simple mask. This operates on the basis that Band 7 is totally absorbed by water, providing a distinct separation." Jupp et al. comment on the difficulty of separating water from steep hillsides with a western aspect or from cloud shadows. The technique suggested by Jupp et al. is to identify deep water visually and then find the minimum values in all the bands and use these values to define the mask. Cloud shadows and steep mountain sides are also identified visually and digitized out of the image manually. Moller-Jensen (1990) classifies water by the following simple rule, using Thematic Mapper data: Band 4 (infrared, wavelength 0.76 to 0.9 pm) < 45 in value (digital number), and Band 5 (infrared, wavelength 1.55 to 1.75 km) < 35 in value. He claims, using this method, that "only a few non-water pixels are misclassified." The cursory treatment given by these, and other, workers to the segmentation of water implies that the problem is straightforward to solve. Under supervision, this is so in many instances, using the methods already described. However, even visual classification can at times be difficult. For example, a small crater lake on the western side of a mountain ridge is virtually impossible to distinguish from deep shadow. One of the images used in this study was deliberately chosen to include such a lake, Lake Euramo in the Danbulla State Forest, Atherton. Automatic Segmentation of Water One objective of this project is to establish some broad principles which can be used for the fully automatic segmentation of water in MSS images which have not been pre-processed, or even de-striped. The images we used have only been radiometrically corrected by the supplier, The Australian Centre for Remote Sensing (ACRES). Most of the commonplace techniques for image interpretation rely on the presence of a human analyst. Methods such as contrast enhancement, histogram equalization, and low pass filtering are representative of those used to assist visual interpretation (Campbell, 1987; Richards, 1986). For fully automatic processing, a change of paradigm is required, with less emphasis on traditional methods and more emphasis on artificial intelligence; hence, the use of rules in this study. The design criteria for the algorithms developed in this research are that the processing of the images should have no human guidance and that the images should not be prePhotogrammetric Engineering & Remote Sensing, Vol. 63, No. 5 , May 1997, pp. 485491. 0099-1112/97/6305-485$3.00/0 O 1997 American Society for Photogrammetry and Remote Sensing processed except for the radiometric correction as outlined previously. The algorithm should not classify any pixel erroneously; i.e., it is acceptable for the software to fail to classify any pixel, but false positives are unacceptable. Any unclassified pixels can be reconsidered later in the application of the algorithm. A misclassified pixel may give rise to incorrect knowledge about an image which could lead to false conclusions. Thus, the Jupp et al. and the Moller-Jensen methods described in the introduction are unacceptable under the last criterion. Another important criterion is that the algorithm must be general enough to identify water under varied conditions as discussed below. Factors Which Cause Varied Reflectances from Water The light reflected from water in an image can vary in intensity in several ways. The change of sun angle on swell, and whitecaps in choppy water result in specular reflection. The water may be turbid, or may contain vegetable matter such as algae or weed. The water may be shallow, giving rise to bottom reflectance (Campbell, 1987, Chapter 14; Lyzenga, 1981). The data values (digital numbers) are also affected by the atmosphere, and the algorithm must cope with variations in atmospheric path radiance and scattered radiance from neighboring pixels. Variability caused by atmospheric radiance has been thoroughly discussed in the literature (Dave et al., 1980; Switzer et al., 1981; Kowalick et al., 1983; Crippen, 1987; de Haan et al., 19911. The algorithm must accommodate noise caused by sensor gain variation (band striping) and quantization noise. Quantization noise is a result of converting a continuous or analog intensity value into a digital value, and the error is increased by expanding the range from the six bits captured by the camera to the eight bits of the data sold by ACRES. The error is increased again non-linearly by the radiometric correction algorithm employed by ACRES. Quantization noise can be greater than 50 percent for the low values of the infrared wavelengths for water, and it is particularly frustrating for developers of automatic algorithms. The algorithm must also accommodate the difficulty of mixed pixels (mixels), both homogeneous (marshland) and edge mixels (shorelines). The author takes the view that the problem of mixels is being attacked in a different study, and classifying mixels, although an important consideration, is beyond the scope of this paper. The Variance Filter Algorithm A useful aspect of water (at least in oceans, wide rivers, and lakes) is that it is uniform and relatively flat, although some local variation in the brightness of water pixels is caused by all of the properties discussed in the previous section. It is the uniformity of water bodies that is exploited by the variance filter. The variance filter provides a measure of local homogeneity in an image. It can also be regarded as a nonlinear non-directional edge detector. The filter emphasizes sudden changes in image brightness without any directional bias and is successful at identifying shorelines. High variance filter values at the shoreline are useful for limiting the region expansion algorithms (described in a later section). Mountain ridges cause very high variance filter values, and this fact is exploited in order to differentiate between mountain shadow and water. The use of a texture channel in Landsat image interpretation is not new. The variance filter is one of a suite of filters based on straightforward statistical measures; mean, median, mode, variance, dispersion, etc. This set of filters, the so-called histogram filters, are used in many image processing applications. The variance filter has been mentioned briefly by Jain (1989, p. 344) as follows: "Variance can be used to measure local activity in the amplitudes." A similar type of filter has been used by Jupp et al. (1985) in the form of a texture channel previously proposed in the work of Haralick and Shanmugam (1974). The texture algorithm is the computed local root-mean-square (RMS) between the center pixel of a box and all the other eight pixels in the box. In the work of Jupp et al., a 3by 3-pixel box was used. Jupp et al. computed the texture only for Band 4 (MSS data] because the green band has the greatest water penetration. Haralick and Shanmugam (1974) specifically include variance in the appendix of their paper. The paper discusses the use of texture for automatic land-use classification, mentioning that water bodies display considerably more homogeneity than does grassland for example. There have been other recent advances, in image classification, proposed by the statistical analysis community, where local homogeneity is used as one property in inferring class. These methods often rely on Bayesian computation and Markov random fields to represent local characteristics in an image (Besag, 1986; Smith and Roberts, 1993). The problems associated with Bayesian computation for Landsat image classification have been well documented in the literature (Skidmore and Turner, 1988). Skidmore and Turner, in their proposed non-parametric supervised classifier, have used a lookup table for modeling non-parametric search spaces. In a closely related approach, the use of rules in this project is a simple and effective way of modeling the nonlinear and non-Gaussian search space of wind-ruffled. turbid. or shallow water. One purpose of this paper is to show that the variance filter is a straightforward way of using local homogeneity for image classification. The use of region-growing algorithms (described later] achieve similar ends to Besag's Markov random fields. The variance filter algorithm consists of replacing a central pixel value with the variance of a specified set of pixel values surrounding it. The set need not be a square set of n by n pixels, where n is an odd number. Indeed, the set can be any specified group. The variance of such a set is given as where N is the number of pixels in the set, x, is the value of pixel i, and x is the mean of pixel values in the set. It makes little difference whether or not the unbiassed estimator definition of the variance is used. For a 5 by 5 set of pixels, N is 25 throughout, so the question of whether to use N o r N 1 is merely a question of scaling factor. The computation is a two-pass one, because the mean of the set must be calculated on the first pass and the variance on the second. Experiments were carried out on individual bands in order to determine an appropriate box size. A 3 by 3 variance filter on Band 7 was very successful in highlighting shorelines but did not smooth noise sufficiently. In particular, some uniform patches of cloud shadow were still indistinguishable from water. A 7 by 7 box was found to smooth edges too much so that some shorelines lost some definition. A 5 by 5 box, with the target pixel in the center, was found to give an optimum balance between definition and smoothing. The Modulus Image The variance filter can be used on any raster image, such as an individual band or any linear or non-linear combination of bands (one of the principle components for example). Very many variations on the algorithm are therefore possible.