Philip J. Howarth

A Comparison of Spatial Feature Extraction Algorithms for Land-Use Classification with SPOT HRV Data

Abstract A large number of spatial feature extraction methods were developed during the past 20 years. The effectiveness of each method has been assessed in different studies using different data. However, there have been few application-oriented studies made to evaluate the relative powers of these methods in a particular environment. In this study, three spatial feature extraction methods have been compared in the land-use classification of the SPOT HRV multispectral data at the rural-urban fringe of Metropolitan Toronto. The first two methods are the well-known gray level co-occurrence matrix (GLCM) and the simple statistical transformation (SST). The third method is the texture spectrum (TS), which was developed recently. Twenty-seven spatial features were derived from the SPOT HRV Band 3 image using these methods. Each of these features or a combination of two of these features were used in combination with the three spectral images in the classification of 10 land-use classes. Results indicated that some spatial features derived using the GLCM and the SST methods can largely improve the classification accuracies obtained by the use of the spectral images only. In addition, average transformed divergence was found to be ineffective in selecting optimal spatial features for land-use classification.

Evaluation Of The Grey-level Co-occurrence Matrix Method For Land-cover Classification Using Spot Imagery

Absfruct-Nine cover types have been classified using a textural/ spectral approach. The texture analysis is based on the grey-level cooccurrence matrix method. Texture features are created from a SPOT near-infrared image using four texture indices, seven window sizes, and two quantization levels. A supervised classification based on the maximum-likelihood algorithm is applied to the three SPOT multispectral bands combined with each texture image individually and to the three bands combined with all four texture images. Classification accuracy is measured by Kappa coefficients calculated from confusion matrices. A factor analysis, based on principal components, is performed to evaluate the contribution to the classification accuracy of each variable involved in the creation of the texture features. The addition of texture features provides a significant improvement in the classification accuracy of each cover type when compared with the results obtained from the multispectral analysis alone. The window size accounts for 90% of the classification variability, 7% is explained by the statistics used as texture measures, and only 3% by the quantization level. There is a window size that optimizes the discrimination of each cover type.

Hyperspectral remote sensing for estimating biophysical parameters of forest ecosystems

Remote sensing has demonstrated wide applicability in the area of estimating and mapping forest physical and structural features. Focus in recent years has been directed towards measuring the biophysical/physiological character of forest ecosystems in order to estimate and predict forest ecosystem health and sustainability. The following reviews the relationship between forest condition and reflectance; remote-sensing measurements (and derivatives) that provide biophysical/physiological information; and the potential of hyperspectral sensors in the measurement of these parameters.

Remote sensing and the measurement of geographical entities in a forested environment. 1. The scale and spatial aggregation problem

Abstract The hypothesis tested in this study was that remote sensing constitutes a particular case of an arbitrary uniform spatial sampling grid used to obtain measurements about geographical entities that induces the scale and aggregation effect responsible for haphazard analysis results. The main objective was to evaluate the impact of measurement scale and spatial aggregation on the information content and classification accuracies of airborne MEIS-II data acquired over a midlatitude temperate forested environment. The original MEIS-II data were resampled to four spatial resolutions, namely 5 m, 10 m, 20 m, and 30 m. Forest classes were established according to three progressive levels of spatial aggregation. Descriptive statistics (Wald-Wolfowitz runs test, mean and variance) were calculated on transects of pixels representing each forest class delineated on the images at every spatial resolution. A maximum-likelihood classification was also performed for each combination of spatial resolution and aggregation level. The results reveal that, except for the mean, changing the measurement scale and the aggregation level of the classes greatly affects the values of the descriptive statistics. The Z value of the Wald-Wolfowitz runs test decreases with decreasing spatial resolution. The effect is more pronounced when the classes are progressively aggregated. For most classes, the variance decreases with the decrease of spatial resolution. In such cases, the impact of changing the measurement scale is greater than the change of aggregation level. Per-class accuracies are also considerably modified depending on the measurement scale and the aggregation level. Within a particular aggregation level, some classes are better classified at fine spatial resolutions, while others require coarser spatial resolutions. Three major conclusions can be stated from these results: 1) The information content of remote sensing images is dependent on the measurement scale determined by the spatial resolution of the sensor; 2) neglecting the scale and aggregation level when classifying remote sensing images can produce haphazard results having little correspondence with the objects of the scene; and 3) there is no unique spatial resolution appropriate for the detection and discrimination of all geographical entities composing a complex natural scene such as a forested environment. These conclusions provide a theoretical foundation from which original solutions to the problem of appropriate scales of measurement for geographical entities can be experimented. Logically, there exists an optimal spatial resolution for each entity of interest, corresponding to its intrinsic spatial and spectral characteristics.

High Spatial Resolution Remote Sensing Data for Forest Ecosystem Classification: An Examination of Spatial Scale

Detailed forest ecosystem classifications have been developed for large regions of northern Ontario, Canada. These ecosystem classifications provide tools for ecosystem management that constitute part of a larger goal of integrated management of forest ecosystems for long-term sustainability. These classification systems provide detailed stand-level characterization of forest ecosystems at a local level. However, for ecological approaches to forest management to become widely accepted by forest managers, and for these tools to be widely used, methods must be developed to characterize and map or model ecosystem classes at landscape scales for large regions. In this study, the site-specific Northwestern Ontario Forest Ecosystem Classification (NWO FEC) was adapted to provide a landscape-scale (1:20 000) forest ecosystem classification for the Rinker Lake Study Area located in the boreal forest north of Thunder Bay, Ontario. High spatial resolution remote sensing data were collected using the Compact Airborne Spectrographic Imager (CASI) and analyzed using geostatistical techniques to obtain an understanding of the nature of the spatial dependence of spectral reflectance for selected forest ecosystems at high spatial resolutions. Based on these analyses it was determined that an optimal size of support for characterizing forest ecosystems (i.e., optimal spatial resolution), as estimated by the mean ranges of a series of experimental variograms, differs based on (i) wavelength, (ii) forest ecosystem class, and (iii) mean maximum canopy diameter (MMCD). In addition, maximum semivariance as estimated from the sills of the experimental variograms increased with density of understory.

Change-detection accuracy assessment using SPOT multispectral imagery of the rural-urban fringe☆

Abstract Studies are being undertaken to determine the capabilities of SPOT multispectral (XS) imagery for providing information on rural-to-urban land conversion. Four procedures are tested. These are visual analysis of two images, visual analysis of a multidate image consisting of two XS2 bands, supervised classification of the two images and supervised classification of the multidate image. Results show that visual analysis of two images and supervised classification of the multidate image provide the best overall classification accuracies at approximately 80%. The best change-detection accuracy of 60% is achieved with supervised classification of the multidate image. Change/no change accuracies are greater than 90%. Although classification accuracies are slightly lower than those achieved with Landsat multispectral scanner imagery, on the SPOT imagery changes in much smaller parcels of land can be observed with greater spatial precision.

Use of the Hough transform in automated lineament

Most lineament mapping is done visually using enhanced images. To automate this procedure partially, algorithms have been developed to imitate some of the visual rules employed by geologists when mapping lineaments. The Hough transform is described, and its capabilities for automated lineament analysis are demonstrated using part of a Landsat TM image of the Canadian Shield near Sudbury, ON, Canada. The results of the automated analyses are compared with the major faults shown on a geologic map of the area and with a visual assessment of an image of the same area undertaken by two remote sensing / geology specialists. The results show that automated interpretation identifies more of the faults than visual interpretation.

Using Landsat-5 thematic mapper and digital elevation data to determine the net radiation field of a Mountain Glacier

Abstract An accurate evaluation of glacier energy balance requires a precise knowledge of surface-cover albedo and emittance. Usually these values are acquired from field measurements. However, the microclimatological effects in mountainous terrain greatly limit the potential for spatial extrapolation of such a set of values. This study overcomes the problem by using the upward radiance values registered on Landsat-5 Thematic Mapper (TM) images and information on the geometry of ter terrain extracted from a digital elevation model (DEM). Studies were performed in the Athabasca Glacier basin, one of the major glacier outlets of the Columbia Icefield in the Canadian Rocky Mountains. The methodology is composed, first, of an automated procedure for the physiographic description of the glacierized basin (cover type, elevation, slope, aspect, horizon profiles, sky-view factor, and the level of enclosing topography) in order to calculate the effects of topography on the radiation balance for each 30 m pixel. Because glacier surface covers usually have high reflective behaviors or distinct emission patterns, this study puts special emphasis on modelling the amount of terrain-reflected or terrain-emitted radiation received on a particular surface. Second, the topographic correction is applied to values of irradiance computed using the LOWTRAN-6 code, a spectrally based radiative transfer model, and atmospheric radiosonde measurements of the vertical temperature, air pressure, and relative humidity profiles. The measured upward radiance values from TM, corrected for path effects, are used to calculate surface-cover albedo and brightness temperature. Third, the daily net radiation field for snow and ice covers is computed to illustrate the contribution of the estimated surface radiative parameters to the glacier snow-and-ice melt analysis. The resultsshow that the irradiance values are accurately modelled within 10% of the radiattion values acquired from a field pyranometer. The calculated glacial surface-cover reflectance and albedo values compare favorably to published information, whereas the measured surface brightness temperature is well within the expected range of values.

Road network detection from SPOT imagery for updating geographical information systems in the rural–urban fringe

Abstract Visual interpretation of high-resolution satellite data has been useful for mapping linear features, such as roads and updating land-use changes. However, it would be beneficial to map new road networks digitally from satellite data to update digital databases using semi-automated techniques. In this paper, an algorithm called Gradient Direction Profile Analysis (GDPA) is used to extract road networks digitally from SPOT High Resolution Visible (HRV) panchromatic data. The roads generated are compared with a visual interpretation of the SPOT HRV multispectral and panchromatic data. The technique is most effective in areas where road development is relatively recent. This is due to the spectral consistency of new road networks. As new road networks are those of most interest to the land manager, this is a useful technique for updating digital road network files within a geographical information system of urban areas.

Land-use classification of SPOT HRV data using a cover-frequency method

Abstract A two-stage classification procedure has been applied to extract land use in a rural-urban fringe environment from SPOT High Resolution Visible (HRV) multi-spectral data. In this procedure, the SPOT HRV data were first classified into twelve land-cover types using a supervised maximum-likelihood classification (MLC). In the second stage, cover frequencies were extracted by moving a pixel window over the land-cover map obtained at the first stage. These cover frequencies were then employed in the classification of 14 land-use classes using a supervised minimum-city-block classifier. Results obtained with the cover-frequency method have been compared with those obtained using the conventional MLC approach. The overall accuracy measured by the Kappa coefficient was 0·462 for the MLC method; it was significantly improved to 0·663 with the cover-frequency method.