Identifying change in a dynamic urban landscape: a neural network approach to map-updating

Identifying change in a dynamic urban landscape: a neural network approach to map-updating

Progress in Planning 61 (2004) 327–348 Identifying change in a dynamic urban landscape: a neural network approach to m...

1MB Sizes 0 Downloads 15 Views

Progress in Planning 61 (2004) 327–348

Identifying change in a dynamic urban landscape: a neural network approach to map-updating Christian Langevina, Douglas A. Stowb a National Imagery and Mapping Agency, US Southern Command, 3511 91st Avenue Miami, FL 33172, USA Department of Geography, San Diego State University, 5500 Campanile Dr, San Diego, CA 98182-4493, USA


Received 7 July 2003; accepted 7 July 2003

0305-9006/$ - see front matter q 2003 Elseiver Ltd. All rights reserved. doi:10.1016/S0305-9006(03)00067-9


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Introduction Change in land cover and land use in urban areas is a dynamic process, such that transitions occur at varying rates and in different locations within the constraints of, or in response to, social, economic and environmental factors. Regional planners and decision makers engaged in a map updating operation for a changing metropolitan or county area require up-to-date information on the nature and impact of urban expansion or transition to more intensive usage (Jensen, 1981; Lillesand and Kiefer, 1994; Ridd and Liu, 1998). Map updating is an intensive task requiring timely and accurate information from multiple sources of data. Change detection and identification with multitemporal remotely sensed and ancillary geographic information systems (GIS) data are informative in that timely assessments can be made of biophysical and human-induced processes on the earth’s surface (Jensen, 1981; Westmoreland and Stow, 1992; Jensen, 1996; Ridd and Liu, 1998). An updated, accurate map can serve as a primary data source for urban modeling (e.g. physical, economic planning, forecast models) when transformed and combined with other qualitative and quantitative sources (Donnay et al., 2001). The primary method of updating land-cover and land-use maps has been through airphoto interpretation. In this process, the full range of human interpretation capabilities can be employed, including the interpreter’s own knowledge of the area. However, it is time consuming, subject to errors of omission and interpreters’ abilities vary greatly. Also, there are limits to the ability of humans to absorb and process large volumes of information (Jensen, 1981; Weber and Coskunoglu, 1990; Westmoreland and Stow, 1992). Computer-assisted methods, however, offer automated or semi-automated approaches to detection and identification of land-cover and land-use change. This chapter examines the use of neural networks (NNs) as a semi-automated tool for identifying change in remotely sensed imagery for efficient mapping of the urbanizing San Diego landscape. Analyses were conducted in the context of a land-cover and land-use classification scheme employed by the San Diego Association of Governments (SANDAG) since 1986. Identification of change is an intensive task for detailed mapping of complex urban scenes. A number of techniques provide from – to information of land-use classes. Postclassification comparison is a common technique that utilizes two (or more) dates of classified data sets, but is highly dependent on the accuracy of the initial classification effort (Jensen, 1996). Spectral change-vector analysis both detects and identifies changes by measuring direction and magnitude in Euclidean-based change space. The image analyst selects a threshold beyond which a pixel is judged to have changed from one class to another (Singh, 1989). Ancillary data have the potential to refine change determinations based on spectral radiometric signatures of land-cover change. Hutchinson (1982), Harris and Ventura (1995) and Jensen (1996) examine the incorporation of spatial and non-spatial ancillary data to supplement spectral-based image classification. Stratification involves classification of sub-scenes of an image that may have spectrally similar classes where errors of commission would otherwise occur. Modification of the classifier includes directly encoding contextual information, such as image texture or a priori probabilities from historical sources to weight class allocation. Post-classification sorting involves refining

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


allocation of confused pixels using criterion or decision rules. Choosing the most applicable and accurate ancillary data is critical so as to avoid introducing more error than already exists (Jensen, 1996). Knowledge-based systems, emulating ‘expert’ knowledge, are a more elaborate approach for incorporating ancillary data, but are difficult to implement because they are time consuming and require careful encoding and representation of knowledge. Neural networks offer a compromise between the complex, data intensive knowledge-based systems and parametric manipulation of imagery and ancillary data (Gong et al., 1996; Graham and Barrett, 1997). Under the rubric of computational intelligence, neural networks can be described as a low-level form of artificial intelligence, which are characterized by an ability to learn and to deal with uncertainty in the course of processing noisy, missing or disparate data (Murray, 1999). NNs consist of simple mathematical processors (i.e. nodes) grouped, layered and linked into a variety of arrangements such as the one depicted in Fig. 1. Since 1989, the most widely researched architecture of NNs is the back-propagation trained multilayer perceptron (MLP) (Kanellopoulos and Wilkinson, 1997), for which Paola and Schowengerdt (1995) provide a comprehensive review. Neural network approaches to change detection and identification have been investigated by Gopal et al. (1994), Gopal and Woodcock (1996), Carpenter et al. (1997), Abuelgasim et al. (1999), Dai and Khorram (1999), and Liu and Lathrop (2002). As with most other research in this area, their focus was on land-cover classification. Seto and Liu (2003) provide one of the first NN change detection studies examining land use transitions using the ARTMAP variant compared with the maximum likelihood approach. Herein lies the lack of NN research intent on urban-specific applications, namely, land-use classification, that this research attempts to address. Barnsley et al. (2001) make the distinction between land-cover and land-use mapping such that the former is a relatively straight forward relationship between cover type and detected reflectance, whereas the latter must account for the spatially complex arrangement of spectrally distinct land-cover types. Economic and cultural factors of land use (i.e. human activity therein) cannot be determined by remote sensing means, hence land-use mapping is an inferred process. The intent of this research was to develop and test the first step of a top-down, semiautomated classification procedure for map updating purposes that will characterize

Fig. 1. A multilayer perceptron classifier illustrating classification of multi-spectral data to produce a thematic map (after Foody, 1995).


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

location, quantity and type of land-use change from previously low intensity usage to more built up uses. A maximum likelihood classification (MLC) was performed to compare with the NN outcome. In subsequent phases of an updating operation, a planner utilizing other imagery, GIS and tabular data could improve the initial effort, such as demonstrated by Westmoreland and Stow (1992). Their ‘on-screen’ interactive method is one approach to refining the identification of land-use change to a more detailed level of classification (e.g. Anderson Level III or more).

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Methods Study site and context The Del Mar Quadrangle within San Diego County, California, USA has been the focus of previous remote sensing research on land-use and land-cover change. This region has experienced rapid growth and conversion for more than two decades (Stow et al., 1990; Westmoreland and Stow, 1992; Stow et al., 1998) (Fig. 2). Thus far, SANDAG has performed two complete image-based land-use updates since the initial mapping effort in 1986 (i.e. 1990 and 1995). This study attempts to move beyond the visual interpretation procedures utilized by SANDAG’s by testing a semi-automated approach to explore the potential of increasing efficiency towards the final map product. The desired end-product for metropolitan and transportation planning engineers is normally a vector-based map of updated land use produced from per-pixel classification. The minimum mapping unit requirement of 1 hectare implies that utilizing H-resolution, 4 m spatial resolution imagery will include a large group of pixels (Strahler et al., 1986). Wharton (1982) and Barnsley and Barr (1996) have addressed this problem in their research by incorporating a second stage of classification that included information from neighboring pixels in an attempt to overcome H-resolution problems of complex urban scenes. Data and tools The first image dataset covering the Del Mar study area includes four frames of colour infrared (CIR) aerial photography acquired in 1990 and scanned at approximately 1 m spatial resolution. The second image dataset is a CIR digital orthophotography quarter quadrangle (DOQQ) produced from aerial photography acquired in 1996 and 1997. The image data are comprised of green, red and nearinfrared wavebands. Reference and ancillary data sets in the form of GIS land-use coverages for 1990 (LU90) and 1995 (LU95) were obtained from SANDAG and explored for their utility in inferring changes in land use. Additionally, the LU90 coverage contained projected land-use (PLU90) attribute data included as an additional data column. Also utilized to validate and edit the reference and ancillary GIS coverages were 1995 10 m spatial resolution, merged SPOT panchromatic and multispectral (Pan-XS) imagery and a 1990, colour-infrared hard copy photograph. Here, ‘reference’ refers to a validation data set of assumed higher accuracy and certainty. However, errors can arise in the development of reference data, which in turn results in uncertainty in accuracy measures of the final map (Stow et al., 1990). The image processing and NN software utilized included ERDAS Imagine, version 8.3.1 and the Stuttgart Neural Network Simulator (SNNS), version 4.1. SNNS is freeware developed at the University of Stuttgart that includes numerous types of NN architectures and training algorithms. A major limitation in using this particular NN software is the need to export and import image data between the format requirements of the software packages, adding to processing costs in terms of time and effort.


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Fig. 2. Merged 1995 SPOT XS/Panchromatic (10 m spatial resolution) scene of the Del Mar Quadrangle study area.

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Pre-Classification image processing Image-to-image registration is an essential step for image-based map updating based on multi-temporal imagery. Image warping based on manual selection of ground control points (Fonseca and Manjunath, 1996) was utilized to individually register 1990 CIR aerial photos to the 1996/1997 DOQ image. Identifiable features in agricultural tracts, street cul-de-sacs in areas under construction, and undeveloped areas of both images were emphasized in the selection of ground control points, as built-up urban areas in the 1990 image were not prevalent. The Del Mar area has irregular and abrupt topography and fitting images with 1 m spatial resolution to a registration model was a challenge. Four scanned CIR aerial photographs from 1990 were registered and mosaicked to the 1996/1997 DOQQ image. Following registration, the multitemporal image pair was spatially aggregated to 4 m spatial resolution to simulate commercial satellite multispectral imagery (e.g. Space Imaging IKONOS and Digital Globe QuickBird) that were not available when this study was initiated. Fifteen ground test points were selected within both the 1 and 4 m images of both dates to obtain RMSE values in three separate trials: (a) dispersed across the entire quadrangle before reducing the study area; (b) across the reduced study area and; (c) in areas of change only (i.e. Under Construction). As is evident in Table 1, there is modest improvement with each step in restricting the GCPs to change areas as opposed to the entire study area. The 1990 and 1996/1997 image data were acquired in September and visual comparison revealed that the photos were acquired at different times of the day as was evident by the extensive shadows and difference in vegetation conditions in the latter imagery. Two data removal steps were performed to reduce confusion during the automated phased of this study. First, the 1997 portion in the southwest was removed to maintain consistency with the map updating cycle (e.g. from-to transition from 1990 to 1996). Next, pixels determined to be predominantly comprised of shadow were removed to simplify data pre-processing and precluded allocation of shadow regions into those land-use classes that have characteristically low reflecting, asphalt type surface materials. The disadvantage of this step is the inability to update some areas and quantify the actual area extent that have or have not undergone land-cover and/or land-use change. Stratification Two general types of land-use change occur in urban environments, in-fill/densification and expansion. The former can be characterized by minor to moderate changes, such as Table 1 RMSE values in m for image-to-image registration Image extent

Entire image Study area only Change areas only

Image resolution One meter pixel

Four meter pixel

21.5 20.9 14.7

20.4 19.5 12.7


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

conversion of office space to light industry. Only a portion of the parcel may have changed, making identification difficult. Expansion involves total conversion of land use and land cover, such as clearing agricultural tracts in preparation for homes. Since this study focuses on semi-automated map updating of land-use changes from multi-temporal image data, the study area was restricted to partially built and non-built areas. The 1996 image was stratified into built and partial-/non-built areas using the 4 m raster LU90 GIS coverage (discussed in the following section), based on the ‘expansion’ categories of Parks, Agriculture, Vacant/Undeveloped and Under Construction. Stratification should enable more effective classification by reducing the image brightness variance that would otherwise be present if the entire image were classified (Hutchinson, 1982). Reference data generation and classification Scheme Inspection of the categorical data provided by the SANDAG GIS database revealed that the LU95 coverage did not correspond with existing land-use conditions in many areas. SANDAG, in planned anticipation of new developments, added many built-land-use polygons over agricultural areas and undeveloped tracts. The LU95 coverage was modified by means of heads-up digitizing with the aid of the 1996 CIR DOQ to coincide with the 1996 image-based scene conditions, hereafter labeled as LU96 and used as reference data. The LU90 coverage was also edited for errors relative to the 4 m image data collected in 1990. Both edited layers were then rasterized to a 4 m resolution grid. The classification scheme was developed by overlaying a subset of the corrected LU90 vector data of the expansion categories on the 1995 merged SPOT XS image of the study area. Seven land-use categories were determined to be logical transitions from one image date to the next, based on the premise of urban expansion as listed in Table 2. To reduce inter-class variance and, hence, classification errors, certain land-use classes were aggregated given their spectral-radiometric characteristics were similar in cover type. Upon further inspection of material types in the 1996 imagery, the most variable land-use class was Agriculture, which included row and field crops, pasture land, modest amounts of bare earth, some small stands of low shrub and a few structures. To reduce the intra-class spectral variability, the calibration data for Agriculture were further divided into: Agriculture—Pasture (non-cultivated or fallow fields), Agriculture—Extensive ‘Green’ (cultivated fields), Agriculture—Extensive ‘Red’ (row crops). The NN architecture (and correspondingly the MLC) were designed with this classification scheme. The sub-divided Table 2 Land-cover and land-use classification change scheme From classes (expansion categories)

To classes

Agriculture Vacant/Undeveloped/Natural Parks Under construction

Residential Industrial/Commercial/Office Schools Commercial Recreation/Landscaped Parks Agriculture Vacant/Undeveloped/Natural Parks Under construction

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Agriculture types were recoded into a single class following classification tests and subsequent assessments of accuracy were based on seven land-use classes.

Rationale for incorporating ancillary information Ancillary data may be helpful in overcoming the spectral and spatial complexity of the urban environment in H-resolution imagery. Three options are available for incorporating ancillary data: (a) preparation of the ancillary data as an additional channel (Gong et al., 1996); (b) utilization of the PLU90 data for generation of a contingency table (e.g. Chi-square analysis) of probable changes in land use; or (c) employment of the PLU90 ancillary data and output signals of the NN classification (e.g. partial and multiple class membership) using a post-classification approach to reclassify confused pixels (Foody, 1995, 1999). Inclusion of ancillary data along with the spectral-radiometric information (option a) offers advantages over the other two approaches. Generating a pseudo-interval scale map of previous land use is relatively simple to implement and has the potential to discriminate overlapping calibration signatures. Option b requires large sample sizes when using contingency tables. Insufficient sample size and modification of the edited LU90 coverage from its original state precluded the valid use of the PLU90 data column of projected land use. Using partial and multiple class membership derived from NN routines in conjunction with GIS map information (option c) is not plausible, again because of the unsuitable nature of the PLU90 data. To implement option a, the 1990 4 m image was subjected to a principal components transformation. The first principal component (PC1) effectively simulates an albedo image of surface materials. The PC1 image was stratified using the LU90 GIS land-use map into three separate images corresponding to the three expansion categories. The mean DN for each

Fig. 3. Pseudo-interval map of average Principal Component 1 (PC1) digital numbers for 1990 expansion land uses.


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

stratified subset was recorded and coded to generate a pseudo-interval data layer, as depicted in Fig. 3. A dataset combining the 1996 image data and ancillary data was developed. NN calibration, validation, classification procedures and accuracy assessment Though the input and output layers of MLPs are predetermined by the number of channels and desired classes, respectively, the highly operator-interactive nature for determining the number of hidden nodes is a disadvantage of this type of NN (Paola and Schowengerdt, 1997; Bischof and Leonardis, 1998). This study adopted the suggested heuristic of Kanellopoulos and Wilkinson (1997) for neural networks with one hidden layer: three hidden nodes per input node. The affects of incorporating ancillary data into the image classification process were examined by building two different NNs, a 4-12-9 NN for the 1996 spectral plus ancillary data and a 3-9-9 NN for classifying the 1996 spectral data only. A common practice for validating the robustness of a NN classification approach is to partition the image data set into three mutually exclusive data sets: calibration, validation, and actual (or the terms training, verification, and actual as utilized by Atkinson and Tatnall, 1997). The validation data, comprising a representative sample of desired classes, is often used to verify that a neural network is performing well and should attain reasonable classification accuracy without over-calibrating. This is followed by classification of the actual data. Following division of the data with GIS-derived polygons in the course of selecting calibration points and validation sites, the data were scaled from the 8-bit, 0 to 255 range to a 0 to 1 range. The individual nodes in the MLP utilize the 0 to 1 scale sigmoid transfer function (Paola and Schowengerdt, 1995; Kanellopoulos and Wilkinson, 1997). The process was reversed after classification trials for accuracy assessments. Calibration data selection involved careful visual assessment of the 1996 spectral plus ancillary image to extract the characteristic signature for each land-use class. Area of interest (AOI) blocks were used to extract 300 or more calibration pixels from two sets of data, one from the 1996 spectral plus ancillary data set to calibrate the 4-12-9 NN and one from the spectral—only data to calibrate the 3-9-9. Table 3 shows the coincident land-use classes when the calibration data were collected from the layer stacked 1990 (pseudointerval data of previous land use) and 1996 (spectral information) data. Accounting for Table 3 Corresponding land-use calibration pixels between imaging dates 1990 ‘From’ LU

1996 ‘To’ LU

Agriculture and Vacant/Undeveloped/Natural Parks and Under Construction Vacant/Undeveloped/Natural Parks and Under Construction Agriculture and Vacant/Undeveloped/Natural Parks and Under Construction Agriculture and Vacant/Undeveloped/Natural Parks Under Construction Agriculture and Vacant/Undeveloped/Natural Parks Vacant/Undeveloped/Natural Parks Agriculture and Vacant/Undeveloped/Natural Parks

Residential Commercial/Office/Industrial Schools Agriculture Commercial Rec. /Landscape Parks Agriculture Vacant/Undeveloped/Natural Parks Under Construction

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Table 4 Calibration error measures of the 3-9-9 and 4-12-9 NN architectures Training Error Measures

R-sqr Adjusted R-sqr MSE RMSE

NN architecture 3-9-9 NN

4-12-9 NN

0.47 0.44 0.50 0.70

0.68 0.67 0.30 0.55

each land-use transition between imaging years ensured that the classifiers identified transition sequences, which were then employed in a single classification. This may be an advantage over the post-classification change identification approach, which can be limited by the accuracy of the individual classifications of each time step (Singh, 1989; Jensen, 1996) as well as being more efficient in having only one image to classify. The NNs were trained to achieve the lowest possible system errors using the backpropagation training algorithm. Each iteration of presenting the calibration data and generating an error between input vector and desired output vector adjusts the connections of weights between the nodes until the system error reaches a minimum. Calibration error measures provided by the SNNS software were the regression coefficient ðR2 Þ; adjusted R2 ; mean square error (MSE), and root mean square error (RMSE), as reported in Table 4. AOI blocks were used to delineate 26 validation sites that represented the spectral and spatial complexities within each land-use type totalling 156,472 pixels. Following classification of the validation sites with the MLC and NN classifiers (with and without the ancillary data), the validation image was classified using the NN only. The LU96 reference data layer was applied to all classifications to obtain agreement measures. The rasterized 1996 GIS reference layer provided a pixel-by-pixel comparison (i.e. wall-to-wall sampling) from which to generate error matrices. This precluded choosing a sampling strategy, confidence intervals and dealing with statistical uncertainty. The overalll, user’s2 and producer’s3 accuracies and Kappa statistic ðKhat Þ4 were calculated for the products resulting from each of the classifications. The overall accuracy and Khat statistic are indicators of classifier performance for the whole classified map, whereas user’s and producer’s measures enable assessment of accuracy for individual classes. 1

Overall accuracy is derived by dividing the total number of correctly classified pixels by the total number of pixels in the sample. 2 User’s accuracy is derived by dividing the total number of correctly classified pixels in a category by the total number of pixels that were actually classified in that category (a measure of commission error). It indicates the probability that pixels classified on the map actually represent that category on the ground (Story and Congalton, 1986; from Jensen (1996: 250)). 3 Producer’s accuracy is derived by dividing the total number of correct pixels in a category by the total number of pixels of that category as derived from the reference data. This measure indicates the probability of a reference pixel being correctly classified and is a measure of omission error (Story and Congalton, 1986; from Jensen (1996: 250)). 4 Kappa analysis is a discrete multivariate statistic used in accuracy assessment. Kappa analysis yields a Khat statistic (an estimate of Kappa) that is a measure of agreement or accuracy (Congalton and Mead, 1983; Rosenfield and Fitzpatrick-Lins, 1986; Jensen, 1996).


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Results and Discussion Classification of validation data set Classification of validation sites using the NN and MLC classifiers enabled classification model calibration and provided an opportunity to validate the efficacy of including the ancillary data. The 4-12-9 NN (spectral plus ancillary data) demonstrated superior classification accuracies in comparison to the 3-9-9 NN (spectral only). The overall accuracies for the 4-12-9 NN and 3-9-9 NN were 65.4% ðKhat ¼ 37:4%Þ and 57.3% ðKhat ¼ 23:4%Þ; respectively (Table 5). On a category-by-category basis, the producer’s accuracy of several categories significantly improved; most notably for Commercial Recreation/Landscaped Parks (95.6 versus 65.3%), Vacant/Undeveloped/Natural Parks (81.0 versus 30.5%), and Under Construction (63.4 versus 21.6%). Modest improvements were observed for Agriculture (88.1 versus 73.6%), Industrial/Commercial/Office (32.9 versus 10.2%) and Schools (33.1 versus 24.0%). The accuracy of Residential class decreased (16.1 versus 26.2%). The greatest benefit of the ancillary data was improvement in the non-built land-use types, with slight improvement or decrease in accuracy in the built land-use types. The MLC trials performed modestly better than the 3-9-9 NN, but significantly lower than that of the 4-12-9 NN (Table 6). With only spectral information, the MLC achieved an overall accuracy of 41.4% ðKhat ¼ 28:1%Þ; whereas with the ancillary data the overall accuracy increased slightly to 42.9% ðKhat ¼ 29:4%Þ: Improvement is evident in the producer’s accuracy for Residential (40.0 versus 25.3%), Industrial/ Commercial/Office (34.6 versus 14.5%), Commercial Recreation/Landscaped Parks (95.0 versus 73.9%) and Agriculture (67.7 versus 51.6%). The most striking results for this MLC comparison is the accuracy of the low Vacant/Undeveloped/Natural Parks category (0.0%), improvements for other categories notwithstanding. The slight improvement in overall and Khat measures is due to the higher agreement in the other categories and relatively equal distribution of pixels per category for the validation data set. Perhaps the most influential condition contributing to the inability of the MLC to classify Vacant/Undeveloped/Natural Parks is the nature of the ancillary data that were utilized. The maximum likelihood classifier makes explicit assumptions about the statistical distribution of the data set. The distribution of the pseudo-interval data designed for this study is multi-modal, in that, out of a possible 256 brightness levels in 8-bit imagery, only three bins contain all the pixels of the ancillary data. As such, this type of ancillary data may be an inappropriate means for artificially generating continuous image data for input for the MLC type of classifier. Neural networks on the other hand are better suited to accept disparate data types (e.g. interval, ratio, etc.), and in this case proved to be of benefit. However, when only the spectral data were input to the two classifiers the differences in accuracy were not significant. At this stage, then, it was determined to discuss only the NN results for the validation data set, as the MLC was deemed inappropriate for use with this particular ancillary data type.

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Table 5 Validation site accuracies for the 3-9-9 NN (top; spectral-only data set) and the 4-12-9 NN (bottom; spectral plus ancillary data set)

Actual image classification The overall accuracy of 75.7% ðKhat ¼ 63:3%Þ for the 4-12-9 NN classification of the spectral plus ancillary data indicate that the method shows significant promise (Table 7; Fig. 4). These accuracies are typically higher than for the validation sites. It must be noted that the ratio of total pixel count between land-use types of the validation sites were similar, whereas the ratio of pixel count in the validation image is


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Table 6 Validation site accuracies for maximum likelihood classifier with spectral information only (top) and with the inclusion of ancillary data (bottom)

largely dominated by the Vacant/Undeveloped/Natural Parks category closely followed by Agriculture, thereby resulting in higher accuracies. Producer’s and user’s accuracy for Agriculture and Vacant/Undeveloped/Natural Parks were the highest at 80.2% and 83.9, and 93.9 and 94.7%, respectively. The producer’s outcome is also acceptable for Commercial Recreation/Landscaped Parks at 83.6%. As with the validation sites trial, the addition of ancillary data did not overcome the spectral complexities of the builtland uses as expressed in H-resolution imagery.

Table 7 Classification error matrix of validation image utilizing the 4-12-9 NN C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348 341


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Fig. 4. Land use classification map of spectral plus ancillary image derived with the 4-12-9 NN classfier.

Spectral-only versus inclusion of previous land-use information Recently constructed or modified land uses in the 1996 image that are spatially distributed among more than one ancillary data class (see Table 4) may have contributed to the low classification accuracies of the built land-use categories. Newly built Residential co-occurs with all three ancillary data classes; Schools co-occurs with Agriculture and Under Construction, and Industrial/Commercial/Office co-occurs with Under Construction. Calibration data were collected in a manner that ensured the inclusion of all land-use types that were evident in 1990. This may have contributed to the spectral intra-class variability inherent in H-resolution imagery set for the built categories. Seasonal differences in plant phenological development and soil moisture between image dates are problematic with respect to accurately identifying real change between

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


Fig. 5. Four sub-sets of imagery and GIS layers demonstrating land cover changes not associated with land change. Area A indicates previous land use (subsets a and b) of Vacant/Undeveloped. Area B indicates previous land use of Agriculture. New surface conditions in 1996 (subset d) and NN classified results of the 1996 spectral plus ancillary test image (subset c) illustrate misclassification of Agriculture as Under/Construction (white portion labelled A in NN classified sub-set) and Schools (B label in light green area in 1996 DOQ subset).

1990 and 1996. This type of misclassification is illustrated in Fig. 5 (i.e. areas A and B in each of the example subsets). Crops harvested just prior to the 1996 acquisition, or newly graded fields from the previously undeveloped lands (area A), result in a large amount of dry exposed soils. The image brightness response is likely to be very similar (bright), hence, Agriculture and Under Construction are easily confused. Another type of confusion occurred between Agriculture and Schools (depicted as area B). Visual comparison between the 1990 and 1996 image subsets shows a change in cover type from fallow field to tilled ground with some natural vegetation re-growth. The ancillary data represents the previous land use as Agriculture, which has not changed in the interim, whereas a subset shows that the 4-12-9 NN classified the area as Schools. The potential for misclassification


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Fig. 6. 1996 image (top) showing Commercial Recreation/Landscaped Parks (medium to bright NIR reflectance) and riparian vegetation (medium NIR reflectance). The 4-12-9 NN classification (bottom) distinguished between these two land use types.

is likely compounded by information on the previous land-use state. In both cases, drastic changes in surface-cover types that did not undergo land-use change proved difficult to classify, a problem that may have to be accounted for in subsequent phases of a mapupdating strategy. Conversely, the pseudo-interval ancillary data of prior land use did benefit the NN classification results in distinguishing land uses, particularly between the Vacant/Undeveloped/Natural Parks and Commercial Recreation/Landscaped Parks. Fig. 6 depicts

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


subsets of the NN-classified map and 1996 image of riparian vegetation and a parcel of the Commercial Recreation/Landscaped Parks category. The riparian vegetation exhibits moderate NIR reflectance, whereas the landscaped vegetation has relatively higher NIR reflectance. Without the input of the pseudo-interval data the riparian areas may have been misclassified as Commercial Recreation/Landscaped Parks.


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Discussion and Conclusions The intent of this research was to develop and test image processing approaches for the first step of a top-down, semi-automated land-use classification procedure for map updating purposes. A method of change identification was examined using artificial neural networks and simulated high-resolution, multispectral imagery of 1996 that incorporated land-use information of 1990. When ancillary data were classified as an additional channel with the spectral information, the NN performed significantly better than the MLC. The pseudo-interval ancillary data proved to be an inappropriate input for the MLC; conversely, input of disparate data types is a characteristic strength of NNs as demonstrated in this study. However, performance of the NN classifier of spectral-only data was no better than the MLC. Research of other NN types for urban land-use change identification applications should include fuzzy ARMAP, which has been demonstrated to be more robust than the MLP type NN (Gopal and Fischer, 1996; Carpenter et al., 1997). Further ARTMAP studies such as that conducted by Seto and Liu (2003) will be necessary. The strength of this type of NN lies in its ability to classify multi-temporal data while still retaining ‘memory’ of calibration data. This is a characteristic shortcoming for the MLP type NN, which has to be trained for each new data set to account for new scene conditions. Other researchers have tested methods for the post-classification stage by which to update land-use maps produced by semi-automated means. Wharton (1982) used a kernel-based approach that calculates the frequency of land-use types, and Barnsley and Barr’s (1996) SPAtial Reclassification Kernel (SPARK) includes the spatial arrangement (edge and vertex adjacencies) of class frequency, by which the land use of the center pixel is inferred. Barnsley et al. (2001) extended this line of investigation with knowledge-based texture and structural pattern recognition approaches. Research by Foody (1995, 1996, 1999) and Moody et al. (1996) suggests that preservation and use of the outputs of fuzzy classifiers are desirable in that the primary and secondary land-use labels correspond to the two highest NN outputs. The difficulties of classifying complex urban landscapes may be addressed with these approaches before passing on the classified image to more detailed, manually intensive phases of the map updating scheme. Steps to improve the initial map-updating products in the beginning phases, such as the approach tested in this research, may help to reduce overall efforts and increase timeliness in the latter phases of generating accurate GIS-based map layers. QuickBird (Digital Globe Inc.) (0.6 m pan and 2.4 m multispectral) and IKONOS (Space Imaging Inc.) (1 m pan and 4 m multispectral) imagery has become increasingly available in the past few years with other platforms to be launched in the near future. If remotely sensed data are to be incorporated in the every day practice of urban planners, further research will be needed to ascertain effective spatial resolutions for semiautomated or automated techniques, and algorithms that will generate meaningful measurements and maps of rapidly urbanizing areas such as southern California. Also, the availability of new and hybridized approaches and techniques, and new software tools will need to be developed and transferred from the experimental to the easily understood and practical (Donnay et al., 2001).

C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348


References Abuelgasim, A.A., Ross, W.D., Gopal, S., Woodcock, C.E., 1999. Change detection using adaptive fuzzy neural networks: environmental damage assessment after the Gulf War. Remote Sensing of Environment 70 (2), 208–223. Atkinson, P.M., Tatnall, A.R.L., 1997. Neural networks in remote sensing. International Journal of Remote Sensing 18 (4), 699–709. Barnsley, M.J., Barr, S.L., 1996. Inferring urban land use from satellite sensor images using kernel-based spatial reclassification. Photogrammetric Engineering and Remote Sensing 62 (8), 949–958. Barnsley, M.J., Møller-Jensen, L., Barr, S.L., 2001. Inferring urban land use by spatial and structural pattern recognition. In: Donnay, J., Barnsley, M.J., Longley, P.A. (Eds.), Remote Sensing and Urban Analysis, Taylor and Francis, New York, pp. 115– 144. Bischof, H., Leonardis, A., 1998. Finding optimal neural networks for land use classification. IEEE Transactions on Geoscience and Remote Sensing 36 (1), 337 –341. Carpenter, G.A., Gjaja, M.N., Gopal, S., Woodcock, C.E., 1997. ART neural networks for remote sensing: vegetation classification from Landsat TM and terrain data. IEEE Transactions on Geoscience and Remote Sensing 35 (2), 308–325. Congalton, R.G., Mead, R.A., 1983. A quantitative method to test for consistency and correctness in photointerpretation. Photogrammetric Engineering and Remote Sensing 49 (1), 69–74. Dai, X.L., Khorram, S., 1999. Remotely sensed change detection based on artificial neural networks. Photogrammetric Engineering and Remote Sensing 65 (10), 1187–1194. Donnay, Barnsley, Longley, P.A., 2001. In: Donnay, J., Barnsley, M.J., Longley, P.A. (Eds.), Remote Sensing and Urban Analysis, Taylor and Francis, New York, pp. 3–18. Fonseca, L.M.G., Manjunath, B.S., 1996. Registration techniques for multisensor remotely sensed imagery. Photogrammetric Engineering and Remote Sensing 62 (9), 1049–1056. Foody, G.M., 1995. Land cover classification by an artificial neural network with ancillary information. International Journal of Geographical Information Systems 9 (5), 527–542. Foody, G.M., 1996. Approaches for the production and evaluation of fuzzy land cover classification from remotely-sensed data. International Journal of Remote Sensing 17 (7), 1317–1340. Foody, G.M., 1999. The continuum of classification fuzziness in thematic mapping. Photogrammetric Engineering and Remote Sensing 65 (4), 443 –451. Gong, P., Pu, R., Chen, J., 1996. Mapping ecological land systems and classification uncertainties from digital elevation and forest-cover data using neural networks. Photogrammetric Engineering and Remote Sensing 62 (11), 1249–1260. Gopal, S., Fischer, M., 1996. A comparison of three neural network classifiers for remote sensing classification. Proceedings of IGARSS 1996, April, Lincoln, Nebraska, pp. 787– 789.. Gopal, S., Sklarew, D.M., Lambin, E., 1994. Fuzzy-neural networks in multi-temporal classification of landcover change in the Sahel. Proceedings of the DOSES Workshop on New Tools for Spatial Analysis 1994, Lisbon, Portugal, DOSES, EUROSTAT ECSC-EC-EAEC, Brussels, Luxembourg, pp. 55–68.. Gopal, S., Woodcock, C., 1996. Remote sensing of forest change using artificial neural networks. IEEE Transactions on Geoscience and Remote Sensing 34 (2), 398–403. Graham, D., Barrett, A., 1997. Knowledge-Based Image Processing System, Springer, London, 178 p. Harris, P.M., Ventura, S.J., 1995. The integration of geographic data with remotely sensed imagery to improve classification in an urban area. Photogrammetric Engineering and Remote Sensing 61 (8), 993–998. Hutchinson, C.F., 1982. Techniques for combining Landsat and ancillary data for digital classification improvement. Photogrammetric Engineering and Remote Sensing 48 (1), 123– 130. Jensen, J.R., 1981. Urban change detection mapping using Landsat digital data. The American Cartographer 8 (2), 127–147. Jensen, J.R., 1996. Introductory Digital Image Processing: A Remote Sensing Perspective, Prentice Hall, Upper Saddle River, NJ, 316 p. Kanellopoulos, I., Wilkinson, G.G., 1997. Strategies and best practice for neural network image classification. International Journal of Remote Sensing 18 (4), 711–725.


C. Langevin, D.A. Stow / Progress in Planning 61 (2004) 327–348

Lillesand, T.M., Kiefer, R.W., 1994. Remote Sensing and Image Interpretation, Wiley, New York, 750 p. Liu, X., Lathrop, R.G. Jr., 2002. Urban change detection based on an artificial neural network. International Journal of Remote Sensing 23 (12), 2513–2518. Moody, A., Gopal, S., Strahler, A.H., 1996. Artificial neural network response to mixed pixels in coarseresolution satellite data. Remote Sensing of Environment 58 (3), 329–343. Murray, T., 1999. Computational intelligence techniques in Geography: an introduction. Journal of Geographical Systems 1 (1), 1–2. Paola, J.D., Schowengerdt, R.A., 1995. A review and analysis of backpropagation neural networks for classification of remotely sensed multispectral imagery. International Journal of Remote Sensing 16 (16), 3033–3058. Ridd, M.K., Liu, J., 1998. A comparison of four algorithms for change detection in an urban environment. Remote Sensing of Environment 63 (2), 95–100. Rosenfield, G.H., Fitzpatrick-Lins, K., 1986. A coefficient of agreement as a measure of thematic classification accuracy. Photogrammetric Engineering and Remote Sensing 52 (2), 223–227. Seto, K.C., Liu, W., 2003. Comparing ARTMAP neural network with the maximum-likelihood classifier for detecting urban change. Photogrammetric Engineering and Remote Sensing 69 (9), 981–990. Singh, A., 1989. Digital change detection techniques using remotely-sensed data. International Journal of Remote Sensing 10 (6), 989 –1003. Story, M., Congalton, R., 1986. Accuracy assessment: A user’s Perspective. Photogrammetric Engineering and Remote Sensing 52 (3), 397–399. Stow, D.A., Collins, D., McKinsey, D., 1990. Land use change detection based on multi-date imagery from different satellite sensor systems. Geocarto International 3, 3–12. Stow, D., Coulter, L., Langevin, C., Service, D., Kaiser, J., 1998. Contextual Classification of Changing Land Use Using Multiple Geo-Spatial Data Sources: A Project of the NASA Affiliated Research Center at SDSU. Affiliated Research Center (ARC), San Diego State University, San Diego, CA.. Strahler, A.H., Woodcock, C.E., Smith, J.A., 1986. On the nature of models in remote sensing. Remote Sensing of Environment 20 (1), 121– 139. Weber, E.U., Coskunoglu, O., 1990. Descriptive and prescriptive models of decision-making: Implications for the development of decision aids. IEEE Transactions on Systems, Man and Cybernetics 20 (2), 310 –316. Westmoreland, S., Stow, D.A., 1992. Category identification of change land-use polygons in an integrated image processing/geographic information system. Photogrammetric Engineering and Remote Sensing 58 (11), 1593–1599. Wharton, S.J., 1982. A contextual classification method for recognizing land use patterns in high resolution remotely sensed data. Pattern Recognition 15 (4), 317 –324.