This article was downloaded by: [UNAM Direccion General de Bibliotecas] On: 10 February 2011 Access details: Access

Details: [subscription number 788841385] Publisher Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 3741 Mortimer Street, London W1T 3JH, UK

International Journal of Remote Sensing

Publication details, including instructions for authors and subscription information: http://www.informaworld.com/smpp/title~content=t713722504

Classification of Landsat Thematic Mapper imagery for land cover using neural networks

M. J. Aitkenheada; I. H. Aaldersb a Department of Plant & Soil Science, 23 St Machar Drive, University of Aberdeen, Scotland, UK b The Macaulay Institute, Craigiebuckler, Scotland, UK First published on: 03 January 2008

To cite this Article Aitkenhead, M. J. and Aalders, I. H.(2008) 'Classification of Landsat Thematic Mapper imagery for land

cover using neural networks', International Journal of Remote Sensing, 29: 7, 2075 2084, First published on: 03 January 2008 (iFirst) To link to this Article: DOI: 10.1080/01431160701373739 URL: http://dx.doi.org/10.1080/01431160701373739

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf This article may be used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

International Journal of Remote Sensing Vol. 29, No. 7, 10 April 2008, 20752084

Classification of Landsat Thematic Mapper imagery for land cover using neural networks
M. J. AITKENHEAD{ and I. H. AALDERS{ {Department of Plant & Soil Science, 23 St Machar Drive, University of Aberdeen, Aberdeen AB24 3UU, Scotland, UK {The Macaulay Institute, Craigiebuckler, Aberdeen AB15 8QH, Scotland, UK
Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

(Received 24 August 2006; in final form 27 March 2007 ) Landsat Thematic Mapper (TM) imagery can be used to classify different land cover types based on reflectance and emittance characteristics in seven wavelength bands. Various methods, including NDVI and other simple mathematical transformations, can be used to show strong variations in band intensity ratios from different surfaces. However, the number of land cover classes used is commonly low, preventing a detailed mapping of the region of interest. A neural network trained with the backpropagation method should be able to improve on these simple mathematical calculations by developing complex functions which allow recognition of different land cover or land use types. Landsat imagery of Aberdeen and the surrounding area was used to develop a land cover map highlighting areas of residential, commercial and industrial land use, along with various natural and semi-natural land cover classes. Confusion between specific classes is highlighted by the use of a Kohonen self-organizing map to categorize the Landsat multispectral imagery, resulting in a description of the land cover categories that can actually be distinguished from one another using Landsat TM imagery.

1.

Introduction

Much research has been carried out into the automation of land cover mapping from remotely sensed information. Various methods exist for distinguishing land cover types in imagery, ranging from simple image analysis measurements to more sophisticated structural definitions. In many cases it is not known what spectral characteristics a particular land cover type has when examined using remotely sensed imagery, and whether a land cover type can have more than one combination of spectral intensities. Additionally, it is often difficult with multispectral information to determine the importance of one bands intensity over that of another. A method that automatically determines the spectral characteristics of each particular land cover type and that can apply this information to a large image rapidly and accurately would obviously be useful. One method which shows great potential in this area is that of neural networks. Neural networks can take image information in many forms. For example, Aria et al. (2003) used backpropagation neural networks to identify land cover type, while Dutra et al. (1998) used greyscale image texture as the input for neural network identification of land cover classes from remote sensing data on Amazonia.
*Corresponding author. Email: m.aitkenhead@abdn.ac.uk
International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online # 2008 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/01431160701373739

2076

M. J. Aitkenhead and I. H. Aalders

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

Nepomuceno et al. (2003) applied neural networks to filtered radar data in an attempt to assess land cover in Amazonia. In general, neural networks have been shown to provide superior land cover categorization capabilities to nearest-neighbour or other statistical methods. They are usually applied in one of two ways: as an unsupervised clustering application which automatically identifies clusters (Poth et al. 2001, Zhao et al. 2005) or to identify predetermined land cover classes using supervised training (Foody and Cutler 2006, Aitkenhead and Dyer, in press). Neural networks also work well when applied in parallel with other techniques, or as a first step in image classification. Katartzis et al. (2004) developed a method of detecting limits of minefields using a suite of image analysis techniques, including neural networks. Schwaiger et al. (1996) adopted an evolutionary approach to land cover classification with neural networks, optimizing their networks topology and dynamics to provide the best performance, while Swinnen et al. (2001) succeeded in using neural networks for measuring sub-pixel proportions of land cover types in imagery from the SPOT-VEGETATION satellite. Here, a relatively simple neural network technique is applied, with constraints and using Landsat Thematic Mapper (TM) data to provide a land cover map for the Aberdeen area that may be of interest to urban planners, rural development agencies and other organizations. It is shown that certain classes are relatively easy to identify using this method and with the satellite imagery used, while others tend to be confused with one another. In order to show the extent to which Landsat TM imagery can be confidently applied to land cover mapping without the use of additional information, a Kohonen self-organizing map is applied to the training data. The self-organizing map produces a clustered categorization with specific classes that can be confidently distinguished from one another. 2. Methods and results

The imagery from which the map was derived is a Landsat TM image of Scotland taken on 26 June 1995, with seven bands at wavelengths 0.450.52, 0.520.60, 0.63 0.69, 0.760.90, 1.551.75, 10.4012.50 and 2.082.35 mm. Ground resolution on the image is 25 m pixels, except for band 6, which has a resolution of approximately 120 m. For this band, the imagery was resampled to provide the same resolution as the other bands and was used in the same manner as the others. Band 6 is not commonly used for land cover classification purposes, despite being shown to contain considerable information relating to land cover classes (Southworth 2004). The imagery was calibrated and corrected for atmospheric distortion prior to use. Top left coordinates for the imagery are 368 000, 815 000 and bottom right coordinates are 397 975, 785 025 using the UK Ordnance Survey coordinate system. The image size is 120061200 pixels, or 900 km2. The neural network used is a multi-layer perceptron and is trained using the backpropagation method, a commonly used error-minimization technique. The network topology is 14 : 15 : 15 : N, where each value denotes the number of nodes in successive layers (N is the number of output nodes and corresponds to the number of land cover categories used), and the network is trained for 50 000 training steps. Each layer is fully connected to the one following it. The use of two hidden layers rather than one is sometimes seen as unnecessary for classification purposes. It has been shown that a single hidden layer with a sufficiently large number of nodes will be capable of universal function approximation (Hornik et al. 1989). However, with two hidden layers the number of nodes required for more than simple curve-fitting is

Classification of Landsat TM imagery using neural networks

2077

greatly reduced from that required when using a single layer. For a detailed description of the backpropagation neural network training method implementation, see Aitkenhead et al. (2003). Here, we added to the standard implementation of the backpropagation method the use of a momentum term that directs the error minimization in the direction of recent trends in weight changes, thus helping to avoid local minima and damp oscillations in the connection weight values. The size of the neural network used was selected in order to provide suitable classification while at the same time prevent overfitting of the data. Trial and error with the network size showed that the performance, measured in terms of the accuracy of the resulting maps, increased only very slightly if a larger network architecture was used, while at the same time greatly increasing the processing time required. Training data for the neural network was obtained by examining the Ordnance Survey 1 : 50 000 map of the Aberdeen area and identifying locations where the land cover was known. This required a degree of local knowledge, with 80 training pixels being identified for each land cover type (five locations, each 464 pixels). The number of training data points, while small in comparison to datasets normally used to train neural networks, was kept deliberately small in order to reflect the difficulty that would normally be associated with identifying training data in the field. If the number of training sites is a significant fraction of the total number of pixels being classified, then the cost-effectiveness of the classification overall is reduced. For each training pixel, the seven Landsat band values were used (0255), with the values normalized on the range [0, 1]. An additional seven values were derived for each pixel, these being the proportional values of each band, relative to the summed intensity over all bands for that pixel. These values were used due to a concern that variations in shading over the image could cause different intensity values for two areas with identical land cover type, when the relative albedo for each band was the same between the two areas. For two pixels of the same land cover class of which one is more brightly lit than the other (which may be due to slope angle or shading) the distance between these two pixels in band phase space may be large although the proportional brightness of each band is the same. This Euclidean distance in band phase space may be greater than that between one of those pixels and a pixel from some other land cover class such as water. Adding seven values to the neural network input that contain proportional intensity information was intended therefore to reduce the phase space separation of pixels with the same spectral signature at different overall intensities. Map validation was necessary to get an indication of how well the neural network was classifying land cover. In order to do this, each generated map was sampled randomly to provide 80 pixels for each of the land cover types, or a total of 880 points. The small number of pixels used for validation reflects the above-stated desire to avoid prohibitive amounts of field work, while providing enough information to determine the accuracy of the classification system. Determination of the actual land cover type at each point (according to the categories used here) was then made using a combination of maps, local knowledge and field excursions. The land cover categories originally designated were selected using knowledge about the characteristics of the local landscape, and included the following: lowdensity residential; high-density residential; commercial; low-density industrial; high-density industrial; urban greenspace; arable; forest; water; and natural. This classification was used to produce a land cover map from which visual examination and comparison to existing land cover maps revealed that (a) in

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

2078

M. J. Aitkenhead and I. H. Aalders

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

general, good classification was achieved over broad areas of the region covered (an overall accuracy of 72.3% was achieved) and (b) several small areas were obviously misclassified, for example in the categorization of some areas of forest and natural land as having dense urban land cover, or where the beaches had been categorized as high-density industry. Table 1 gives a confusion matrix for this map, showing where misclassification occurred between land cover classes, and giving the confusion values as percentages of the total number of points for each class. Two reasons were identified for this misclassification tendency. The first was due to poor selection of training areas, resulting in a training dataset that did not accurately reflect the characteristics of the land cover classes used. The areas that had been selected as representative of the land cover classes did in fact contain those classes, but the areas selected in many cases only represented a narrow range of the possible appearances of those classes. For example, all of the forest pixels selected for training were taken from one small portion of the study area, and were representative only of coniferous forestry. The second reason for the misclassification was due to poor selection of land cover categories, resulting in correct classification of areas from the point of view of the network, but not from that of the user. For example, there was a strong tendency of the trained network to identify areas of grassland in rural areas as being urban greenspace, when in fact a better categorization of these pixels is as agricultural grassland. The training sites and land cover categories were reviewed and redefined, with the following land cover definitions being used to generate a new map, using new training data: low-density residential; high-density residential; commercial; lowdensity industrial; high-density industrial; grass; crops; forest; water; natural; and bare ground. There are two main differences in the land cover classifications used between the first and second classification systems applied here. The first was due to a recognition that certain areas of arable land are under grass, which appears similar to urban greenspace. This necessitated a redefinition of arable as crops and grass, with urban greenspace being included in the grass category. The second difference was the inclusion of bare ground, to allow separate classification of sand and bare ploughed fields that were otherwise confused with one another. Another difference introduced at this point is the presence of large areas of unclassified ground. This was due to a decision to accept only those pixel
Table 1. Confusion matrix for initial land cover classification attempt (columns give actual class, rows give predicted class), with the following land cover classification: 1, low-density residential; 2, high-density residential; 3, commercial; 4, low-density industrial; 5, high-density industrial; 6, urban greenspace; 7, arable; 8, forest; 9, water; 10, natural. 1 1 2 3 4 5 6 7 8 9 10 66.2 8.9 6.4 7.0 5.6 3.5 0.8 1.3 0.0 0.3 2 10.8 68.0 9.3 3.2 7.7 0.2 0.5 0.3 0.0 0.0 3 8.3 5.6 69.9 9.5 3.8 1.8 0.1 0.6 0.0 0.5 4 5.3 5.6 3.3 66.0 10.1 3.5 2.0 2.4 0.2 1.5 5 4.3 6.3 4.7 11.3 69.8 2.4 0.8 0.5 0.0 0.1 6 3.0 2.2 2.4 1.9 1.7 69.4 13.0 4.4 0.2 1.8 7 2.2 1.4 1.7 2.4 1.0 8.2 81.8 0.9 0.0 0.5 8 6.3 9.5 1.4 0.6 0.5 0.1 0.2 79.9 0.1 1.5 9 2.6 3.5 2.2 1.8 1.3 1.0 1.1 0.2 84.2 2.0 10 5.0 10.6 1.7 1.9 1.0 2.4 3.2 1.0 0.2 73.0

Classification of Landsat TM imagery using neural networks

2079

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

classifications by the network in which one output node had been a clear and obvious winner, through having an activation value of at least 0.2 more than the second most activated output node. The fact that this resulted in large numbers of unclassified pixels demonstrated that it would be a mistake to adopt a simple winner-takes-all strategy for land cover classification without accepting that because there is a certain proportion of pixels that have spectral signatures similar to more than one land cover category, the probability of misclassification is higher. In table 2, a confusion matrix shows the accuracy of the map once these problem pixels have been removed using a comparison with neighbouring pixels. To classify a problem pixel, the most activated candidate was accepted as long as there was another pixel neighbouring this one that was already classified as the same land cover type. In this way, all but 8622 of 73 491 problem pixels were eliminated. Selecting 100 of these identified problem pixels at random and determining the land cover that actually was present at that point, it was found that the level of accuracy achieved (76.6%) was comparable with the overall accuracy of the image (79.1%). Examination of table 2 showed that good classification appeared to have been achieved, except for the separation of low-density and high-density urban land cover classes. A large proportion of the pixels classified as high-density urban land cover were originally problem pixels, meaning that their classification had not been certain in the first place. This pointed towards an inability of the neural network to distinguish between these two classes using the architecture and training method used. The network was therefore retrained with a different topology of 14 : 30 : 30 : 11, for a total of 500 000 training steps rather than the 50 000 originally used, and with a slower and more accurate learning rate. This more comprehensive training resulted in the more accurate map whose confusion matrix is given in table 3, in which only 46 782 problem pixels were encountered and which had an accuracy (measured as the mean of the accuracy levels achieved over each land cover class) of 84.7%. Three of the land cover classes used, forest, crops and natural land cover (meaning heathland) were not only identified more accurately than this value but predominate within the study region, meaning that the actual proportion of pixels that have been classified correctly within the image is probably higher, approaching 90%.

Table 2. Confusion matrix for improved land cover classification attempt (columns give actual class, rows give predicted class), with the following land cover classification: 1, lowdensity residential; 2, high-density residential; 3, commercial; 4, low-density industrial; 5, high-density industrial; 6, grass; 7, crops; 8, forest; 9, water; 10, natural; 11, bare ground. 1 1 2 3 4 5 6 7 8 9 10 11 70.7 12.8 3.9 4.8 4.4 0.6 0.6 0.9 0.0 0.3 1.0 2 14.2 71.3 4.7 3.1 4.8 0.8 0.3 0.2 0.1 0.1 0.5 3 9.7 8.6 67.0 6.5 4.9 0.9 0.6 0.8 0.1 0.2 0.7 4 9.4 7.7 4.8 70.7 5.1 0.5 0.8 0.5 0.0 0.0 0.6 5 5.5 7.3 6.5 7.4 71.3 0.7 0.3 0.5 0.2 0.0 0.5 6 1.0 1.0 0.8 1.3 1.1 85.0 6.4 1.4 0.1 1.8 0.1 7 0.7 0.3 1.3 1.4 1.0 8.2 82.0 3.2 0.2 1.3 0.5 8 1.3 1.6 0.9 0.6 0.7 3.2 4.9 85.6 0.1 1.1 0.1 9 0.2 0.1 0.3 0.1 0.5 0.8 0.5 0.3 96.7 0.3 0.1 10 0.5 0.7 0.8 0.8 0.3 3.8 4.3 2.4 0.2 85.7 0.6 11 0.9 0.8 1.1 0.6 0.7 2.0 1.8 2.3 2.5 3.5 83.8

2080

M. J. Aitkenhead and I. H. Aalders

Table 3. Confusion matrix for final land cover classification map using smaller learning rate and greater training iterations (columns give actual class, rows give predicted class), with the following land cover classification: 1, low-density residential; 2, high-density residential; 3, commercial; 4, low-density industrial; 5, high-density industrial; 6, grass; 7, crops; 8, forest; 9, water; 10, natural; 11, bare ground. 1 1 2 3 4 5 6 7 8 9 10 11 75.8 10.8 3.2 4.2 3.3 0.6 0.3 0.6 0.1 0.5 0.7 2 10.6 78.1 3.8 2.5 3.2 0.8 0.2 0.2 0.1 0.1 0.5 3 6.6 5.3 79.5 4.1 2.0 0.9 0.3 0.5 0.2 0.0 0.5 4 6.0 6.4 4.2 78.1 3.5 0.5 0.6 0.3 0.0 0.0 0.5 5 4.3 5.5 4.1 5.9 78.2 0.7 0.3 0.3 0.2 0.0 0.5 6 1.3 1.0 0.9 0.8 0.8 88.2 4.4 1.1 0.1 1.3 0.1 7 0.3 0.2 0.6 0.5 0.5 6.1 89.7 1.1 0.2 0.6 0.2 8 0.7 0.9 0.6 0.3 0.5 1.9 3.0 91.3 0.1 0.7 0.1 9 0.2 0.1 0.1 0.1 0.5 0.5 0.2 0.3 97.6 0.2 0.1 10 0.5 0.5 0.5 0.7 0.3 3.4 3.2 1.7 0.5 88.5 0.3 11 0.7 0.8 0.9 0.6 0.5 2.0 1.5 2.2 1.8 2.5 86.6

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

Neural networks do not always converge to the same solution, and it is useful to know whether a large variation in accuracy can be expected from training a network using the above method with the same training data. The backpropagation training method was applied 10 times to a randomly initialized neural network, using the same architecture and training parameters as given above. The accuracy of the network was determined in each case using the same test data, and was found to have a mean value of 85.0% with a standard deviation of 2.3%. This variation in accuracy is considered sufficiently small to provide a reliable land cover mapping method. Table 4 gives the layer size, learning rate, momentum term and number of iterations for each neural network. As the results of training a neural network with the user-selected land cover classification show, there is confusion between certain specific pairs of classes, while for some classes the system is capable of identifying that class nearly perfectly. This raises the problem that regardless of how well the user selects their desired land cover categorization to fit the imagery, or of how accurately they select training pixels to reflect that particular categorization, there will always be some confusion between classes due to spectral similarity. Neural networks are good at discriminating between land cover classes from multispectral imagery, but the natural spread in pixel intensities across a specific class means that there will always be some overlap between the classes in terms of appearance. In addition, different remote sensing platforms will provide different imagery, meaning that at the very
Table 4. Neural network design parameters. Network Inputs Hidden layer 1 nodes Hidden layer 2 nodes Outputs Learning rate Momentum term Training iterations Initial attempt 14 15 15 10 0.05 0.5 50 000 Improved attempt 14 15 15 11 0.05 0.5 50 000 Final attempt 14 15 15 11 0.02 0.5 500 000

Classification of Landsat TM imagery using neural networks

2081

best, it can still only be possible to optimize land cover categorizations for specific platforms. One potential solution to this problem is to use an unsupervised neural network training method to identify clusters that occur naturally within a specific form of imagery. We used the Kohonen self-organizing map in order to identify a set of land cover classes that are spectrally distinct from one another and which can be used to map the entire study region. The Kohonen self-organizing map operates by gradually changing the initially randomized vectors of points in the variable phase space, over several thousand exposures to members of the selected dataset. On exposure to a particular data point, the best matching vector within an array of cells, each containing one vector, is found. Vectors allocated to specific cells are then adjusted according to their distance from this best matching unit, with the adjustment being in the direction of the selected data point vector and decreasing with distance from the best matching unit. The adjustments are also decreased in magnitude over time, with the effect being that the cell vectors move rapidly away from their initial random directions and magnitudes and settle down into spatially organized (within the array) clusters of cells with nearly identical vectors. The end result of this process is an array within which groups of neighbouring cells have approximately the same vector, with clear boundaries between these groups. The trained array can then be used to identify the set of vectors which correspond to classes within the data. The Kohonen selforganizing map can also be used to identify specific land cover classes within remotely-sensed imagery (e.g. Doucette et al. 2001), but they are particularly suited to situations where the clusters identified correspond to specific land cover classes of interest. Training of the Kohonen self-organizing map with Landsat imagery for the study region was carried out by creating a list of the band values for each pixel in the image, and training the network by repeatedly selecting a pixel at random and presenting it to the network. The network was then analysed to determine the spectrally distinct exemplars that it had learned to identify. This process was repeated 10 times in order to determine if the network identified the same classification system every time, or whether there was some variation. We found that each time, approximately the same set of band vectors were identified, with minor variation in values, which led us to conclude that there are specific land cover types in north-east Scotland that are spectrally distinct in Landsat TM imagery. Measuring the vector distance between these class vectors and the satellite imagery, we found that the class vectors closely approximated to the following generic land cover types: water; heath/bog; urban; arable/grass; and bare ground. Selecting training data for these land cover types from the image, using the same methods as described above, and training a new neural network to map the land cover, we found that the land cover can be mapped with an accuracy of 94% if the assumption is made that every pixel in the image must fall into one of these land cover categories. The land cover categorization used here would obviously not be appropriate globally, meaning that the unsupervised clustering method would have to be applied whenever a new region of interest is being studied. 3. Discussion

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

The use of neural networks in classifying land cover from remote sensing imagery, while not a novel concept, definitely has scope for improvement as shown here. Requirements for a successful system include (a) optimal selection of training data,

2082

M. J. Aitkenhead and I. H. Aalders

and (b) determination of land cover classes that can be distinguished using individual pixel spectral signatures. An additional consideration is the level of faith placed in the trained network, in other words how great a distinction has to be made by the network between the winning land cover candidate and that in second place, before the user is satisfied that the network has made a correct identification. An example used here in which relatively large certainty was required of the network still managed to result in a high proportion of the image being classified, but this was after a adjustment of the initial training dataset and land cover legend. In short, neural networks should not be trusted too implicitly to give good land cover maps without a certain level of expertise having gone into the determination of what they are looking for on the ground. Having said that, the land cover map eventually obtained here once the method had been refined had a high level of accuracy with the subjective land cover classes used. Improved levels of accuracy can be achieved by optimizing the land cover classification system used to the imagery that is available. However, this may not always be of use to the user if they are interested in a specific set of land cover classes. Remote sensing imagery, when used to train an unsupervised clustering method prior to interpretation, can provide an accurate broad categorization of the study area that can then be improved upon by the use of additional information. It cannot be expected to provide all of the information that the user requires for mapping purposes. Other remote sensing platforms, such as IKONOS or SPOT, cannot be expected to improve greatly on the level of detail achieved here, except to provide different land cover class optimizations, as the amount of information in their bands is not greatly different from Landsat in terms of quantity or quality. Even if they or other image sources were capable of providing 10 rather than five spectrally distinct land cover classes, this would still leave the user having to either rely on this or try to improve the situation using additional information. Improvement of the basic neural network method would require the addition of an extra technique, as it is felt that neural networks by themselves cannot improve much beyond what is available here. Civco et al. (2002) carried out a comparison of several methods of land cover and land cover change detection, and concluded that no single method can be used to solve all of the problems involved, while Tadesse et al. (2003) showed that an object-recognition approach often worked better than a pixel-by-pixel approach for land cover class recognition. Structural measurements carried out in parallel to analysis of spectral information may prove useful for identifying problem pixels which resist more simple forms of identification. Moriyama et al. (2004) discussed a method of estimating terrain, or land cover, from Synthetic Aperture Radar (SAR). The scattering effects of SAR from different surfaces would prove useful in determining land cover type, as an adjunct to visual and infrared spectral intensity. Shah and Gandhi (2004) discuss the use of textural information in improving accuracy of neural network mapping methods. Textural descriptions contain information about the structure of the location being examined, and can assist methods that use simple greyscale as input. In terms of improving the categorization detail of maps generated with Landsat (or other remotely sensed) imagery, where does this leave us? If we make the assumption that other remote sensing platforms will differ only in the types of land cover classes that are optimal for their imagery specifications, then we are left with the problems of using additional data to increase the level of detail in the mapping system. These include issues of scale, accuracy and the integration of quantitative

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

Classification of Landsat TM imagery using neural networks

2083

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

datasets such as are obtained using remote sensing with other types of data, as well as simply having the problems of additional data acquisition. There are several possible solutions to these problems, although issues of acquisition and data rights will always remain. Combining information derived from quantitative datasets with that obtained from other types of data, such as soil or land cover maps where categories rather than values are used, can be achieved through a harmonization of the data and the use of a common framework. Expert systems methodologies, such as decision trees, can be made to function well with a combination of different data types, while neural networks are more particularly suited to classification problems involving noisy data from disparate data sources (Lek and Guegan 1999). The main unresolved problem that we perceive is that of developing the expert knowledge that will allow one of many possible solutions to be implemented. This knowledge can be acquired through field excursions or through the implementation of existing expert knowledge within some descriptive framework, but it has to be obtained, interpreted and validated before it can be used in the development of a system capable of using it to produce land cover maps. References
AITKENHEAD, M.J. and DYER, R., in press, Improving land cover classification using recognition threshold neural networks. Photogrammetric Engineering & Remote Sensing. AITKENHEAD, M.J., MCDONALD, A.J.S., DAWSON, J.J., COUPER, G., SMART, R.P., BILLETT, M., HOPE, D. and PALMER, S., 2003, A novel method for training neural networks for time-series prediction in environmental systems. Ecological Modelling, 162, pp. 8795. ARIA, E.H., AMINI, J. and SARADJIAN, M.R., 2003, Back propagation neural network for classification of IRS-1D satellite images. In Proceedings, Joint Workshop of ISPRS Working Groups I/2, I/5, IC WG II/IV and EARSeL Special Interest Group: 3D Remote Sensing, High Resolution Mapping from Space 2003, 68 October 2003, Hanover. CIVCO, D.L., HURD, J.D., WILSON, E.H., SONG, M. and ZHANG, Z., 2002, A comparison of land use and land cover change detection methods. In Proceedings, 2002 ASPRS Annual Convention, Washington, DC. DOUCETTE, P., AGOURIS, P., STEFANIDIS, A. and MUSAVI, M., 2001, Self-organized clustering for road extraction in classified imagery. Journal of Photogrammetry & Remote Sensing, 55, pp. 347358. DUTRA, L.V., HUBER, R. and HERNANDEZ, P., 1998, Primary forest and land cover contextual classification using JERS-1 data in Amazonia, Brazil. In Proceedings of International Geoscience and Remote Sensing Symposium, Seattle, WA, USA, July 1998. FOODY, G.M. and CUTLER, M.E.J., 2006, Mapping the species richness and composition of tropical forests from remotely sensed data with neural networks. Ecological Modelling, 195, pp. 3742. HORNIK, K., STINCHCOMBE, M. and WHITE, H., 1989, Multilayer feedforward networks are universal approximators. Neural Networks, 2, pp. 359366. KATARTZIS, A., VANHAMEL, I. and SAHLI, H., 2004, Remote sensing minefield area reduction: model-based approaches for the extraction of minefield indicators. ESA-EUSC 2004: Theory and Applications of Knowledge Driven Image Information Mining, with focus on Earth Observation, Madrid, Spain, 2004. LEK, S. and GUEGAN, J.F., 1999, Artificial neural networks as a tool in ecological modelling, an introduction. Ecological Modelling, 120, pp. 6573. MORIYAMA, T., URATSUKA, S., UMEHARA, T., MAENO, H., SATAKE, M., NADAI, A. and NAKAMURA, K., 2004, Feature extraction of urban areas based on polarimetric

2084

Classification of Landsat TM imagery using neural networks

Downloaded By: [UNAM Direccion General de Bibliotecas] At: 11:09 10 February 2011

decomposition. In Proceedings, 5th European Conference on Synthetic Aperture Radar, 2527 May 2004, Ulm, Germany, pp. 435438. NEPOMUCENO, A., SILVA, N., FREITAS, C., VALERIANO, D., SANTA ROSA, A., DUTRA, L. and SANTOS, J., 2003, P-band radar data classification by neural network for Amazonian land cover assessment. In Proceedings, Geoscience and Remote Sensing Symposium, 2003, pp. 26202622. POTH, A., KLAUS, D., VO, M. and STEIN, G., 2001, Optimization at multi-spectral land cover classification with fuzzy clustering and the Kohonen feature map. International Journal of Remote Sensing, 22, pp. 14231439. SCHWAIGER, R., MAYER, H.A. and HUBER, R., 1996, Evolution of low complexity artificial neural networks for land cover classification from remote sensing data. In Seggau: Pattern Recognition 1996, Schriftenreihe der Osterr. Computer Gesellschaft Band 90, A. Pinz, (Ed.) (Oldenbourg: Vienna), pp. 7586. SHAH, S.K. and GANDHI, V., 2004, Image classification based on textural features using artificial neural network (ANN). Electronics and Telecom Engineering, 84, pp. 7277. SOUTHWORTH, J., 2004, An assessment of Landsat TM band 6 thermal data for analysing land cover in tropical dry forest regions. International Journal of Remote Sensing, 25, pp. 689706. SWINNEN, E., EERENS, H. and LISSENS, G., 2001, Sub-pixel land cover classification with SPOT-VEGETATION imagery. In Proceedings of the International Geoscience and Remote Sensing Symposium 2001, 913 July, Sydney, Australia. TADESSE, W., COLEMAN, T.L. and TSEGAYE, T.D., 2003, Improvement of land use and land cover classification of an urban area using image segmentation from Landsat ETM data. In Proceedings of the 30th International Symposium on Remote Sensing of the Environment, 1014 November, 2003, Honolulu, Hawaii. ZHAO, S., KE, C., DONG, X., LI, J. and FENG, X., 2005, Landsat TM multi-spectral classification using support vector machine method in low-hill areas. Proceedings of SPIE The International Society for Optical Engineering, 5657, pp. 8289.