The paper presents accuracy comparison of sub-pixel classification based on medium resolution Landsat data and high resolution RapidEye satellite images, performed using machine learning algorithms built on decision and regression trees method (C.5.0 and Cubist). The research was conducted in southern Poland for the catchment of the Dobczyce Reservoir. The aim of the study was to obtain image of percentage impervious surface coverage and assess which data sets can be more applicable for the purpose of impervious surface coverage estimation.
Imperviousness index map generation was a two-stage procedure. The first step was classification, which divided the study area into two categories: a) completely permeable (imperviousness index less than 1%) and b) fully or partially impervious areas. For pixels classified as impervious, the percentage of impervious surface coverage within a single pixel was estimated. Decision and regression trees models construction was done based on training data set derived from Landsat TM pixels as well as for fragments of RapidEye images corresponding to the same Landsat TM training pixels. In order to obtain imperviousness index maps with the minimum possible error we did the estimation of models accuracy based on the results of cross-validation. The approaches guaranteeing the lowest means errors in terms of training set using C5.0 and Cubist algorithm for Landsat and RapidEye images were selected.
Accuracy of the final imperviousness index maps was checked based on validation data sets. The root mean square error of determination of the percentage of the impervious surfaces within a single Landsat pixel was 9.9% for C.5.0/Cubist method. However, the root mean square error specified for RapidEye test data was 7.2%. The study has shown that better results of two-stage imperiousness index map estimation using RapidEye satellite images can be obtained.
The aim of our research was to evaluate the applicability of textural measures for sub-pixel impervious surfaces estimation using Landsat TM images based on machine learning algorithms. We put the particular focus on determining usefulness of five textural features groups in respect to pixel- and sub-pixel level. However, the two-stage approach to impervious surfaces coverage estimation was also tested. We compared the accuracy of impervious surfaces estimation using spectral bands only with results of imperviousness index estimation based on extended classification features sets (spectral band values supplemented with measures derived from various textural characteristics groups).
Impervious surfaces coverage estimation was done using decision and regression trees based on C5.0 and Cubist algorithms. At the stage of classification the research area was divided into two categories: i) completely permeable (imperviousness index less than 1%) and ii) fully or partially impervious areas. At the stage of sub-pixel classification evaluation of percentage impervious surfaces coverage within single pixel was done. Based on the results of cross-validation, we selected the approaches guaranteeing the lowest means errors in terms of training set. Accuracy of the imperviousness index estimation was checked based on validation data set. The average error of hard classification using spectral features only was 6.5% and about 4.4% for spectral features combining with absolute gradient-based characteristics. The root mean square error (RMSE) of determination of the percentage impervious surfaces coverage within a single pixel was equal to 9.46% for the best tested classification features sets. The two-stage procedure was utilized for the primary approach involving spectral bands as the classification features set and for the approach guaranteeing the best accuracy for classification and regression stage.
The results have shown that inclusion of textural measures into classification features can improve the estimation of imperviousness based on Landsat imagery. However, it seems that in our study this is mainly due higher accuracy of hard classification used for masking out the completely permeable pixels.
The paper presents accuracy comparison of subpixel classification based on medium resolution Landsat images, performed using machine learning algorithms built on decision and regression trees method (C.5.0/Cubist and Random Forest) and artificial neural networks. The aim of the study was to obtain the pattern of percentage impervious surface coverage, valid for the period of 2009-2010. Imperviousness index map generation was a two-stage procedure. The first step was classification, which divided the study area into categories: i) completely permeable (imperviousness index less than 1%) and ii) fully or partially impervious areas. For pixels classified as impervious, the percentage of impervious surface coverage in pixel area was estimated. The root mean square errors (RMS) of determination of the percentage of the impervious surfaces within a single pixel were 11.0% for C.5.0/Cubist method, 11.3% for Random Forest method and 12.6% using artificial neural networks. The introduction of the initial hard classification into completely permeable areas (with imperviousness index <1%) and impervious areas, allowed to improve the accuracy of imperviousness index estimation on poorly urbanized areas covering large areas of the Dobczyce Reservoir catchment. The effect is also visible on final imperviousness index maps.
In this paper we present the results of research carried out to assess the usefulness of wavelet-based measures of image texture for classification of panchromatic VHR satellite image content. The study is based on images obtained from EROS-A satellite. Wavelet-based features are calculated according to two approaches. In first one the wavelet energy is calculated for each components from every level of decomposition using Haar wavelet. In second one the variance and kurtosis are calculated as mean values of detail components with filters belonging to the D, LA, MB groups of various lengths. The results indicate that both approaches are useful and complement one another. Among the most useful wavelet-based features are present not only those calculated with short or long filters, but also with the filters of intermediate length. Usage of filters of different type and length as well as different statistical parameters (variance, kurtosis) calculated as means for each decomposition level improved the discriminative properties of the feature vector consisted initially of wavelet energies of each component.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.