OBJECT-BASED CLASSIFICATION USING MULTI-RESOLUTION IMAGE FUSION

Abstract

Multisource remote sensing data has recently gained much greater attention of researchers for urban land-cover classification because it is increasingly being realized that the complimentary characteristics of different types of remote sensing data can significantly enhance and aid the identification of natural and man-made objects in an urban landscape. However while on one hand, it improves the classification accuracy, on the other, it also increases the data volume including noise, redundant information and uncertainty between the datasets. Therefore, it is essential to extract selected input features and combine them from the multisource data to achieve the highest possible classification accuracy. The otherchallenging tasks while dealing with multisource data are the development of the data processing and classification techniques to exploit the advantages of multisource data sets. The objective of this thesis is to improve the urban land-cover classification process by using multisource remote sensing data and exploring various spectral and spatial features with recent image processing and classification algorithms and techniques for improvement in classification accuracy of urban land cover objects. This work proposes a methodological approach for classifying natural (i.e., vegetation, trees and soil) and man-made (i.e., buildings and roads) objects using multisource datasets (i.e., Long Wave Infrared (LWIR) Hyperspectral and High Spatial Resolution Visible RGB data). The spatial, spectral and spatial-structural features, such as textural (i.e., contrast and homogeneity), Normalized Difference Vegetation Index (NDVI), Morphological Building Index (MBI) are extracted and incorporated on the connected components of each class under the category of natural and man-made objects. The feature knowledge set formed consisting of different domains of spectral and spatial attributes are trained and tested using one-against-one Support Vector Machine (SVM) classifier network. The decisions of the classifier network are finalized using fusion technique i.e., majority vote contributed by each feature set. The results vi obtained from the classification approach using decision level fusion show that majority vote by NDVI feature in SVM classifier network has contributed in achieving good confidence measures for classes, building (100%), vegetation (91.6%), soil (91%) and road (79.3%). From these results, it may be inferred that NDVI feature gives better results for almost all classes. Since, NDVI is calculated using mean spectral information of thermal bands from hyperspectral data and red bands from VIS RGB data, therefore, the mean spectral information from the thermal hyperspectral bands can be explored for studying the natural and man-made objects in urban land-cover. Also, the feature level fusion of multisource datasets i.e., the fusion of long wave thermal bands from hyperspectral data and red bands from very high resolution visible RGB data has resulted in achieving good classification accuracy of objects in an urban area. A decision levelstrategy has also been derived using the combination of textural, NDVI and MBI features using weighted majority voting fusion technique to enhance the classification results of natural and man-made objects. The comparative analysis of decision level fusion of each feature set and the combination of all features reveal that the overall confidence measure of the classified objects has improved from 57.1% to 98.2% (for all classes except for building). Hence, it is seen that the proposed weighted majority voting strategy of combined spectral and spatial features derived from multisensory data improves the classification accuracy of urban land-cover objects. The capabilities of non-parametric classifiers, such as Artificial Neural Network (ANN) and Support Vector Machine (SVM), have been investigated using LWIR and Visible RGB data with the methodological approach derived using spectral and spatial features and decision level fusion strategy. The results have shown that SVM classifier network gives better generalisation results than ANN for identification of natural and man-made objects. The vii statistics obtained show that in comparison to neural networks, SVM requires less training data for a decent performance. This research also explores the benefit of using wide range hyperspectral data i.e., Airborne Visible/Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG). The spectral channels of hyperspectral data ranging from 325-nm to 2500-nm have been investigated for extracting the useful wavelengths for forming the spectral indices which in turn can be used for identifying the components /constituents of natural and man-made objects of an urban area. The Optimum Index Factor (OIF) is computed by evaluating different standard deviation and correlation values. The wavelength values for which the best OIF is obtained are selected to form the spectral indices. Various spectral indices (namely, Normalised Difference Vegetation Index (NDVI), Infrared Percentage Vegetation Index (IPVI), Difference Vegetation Index (DVI)-wavelengths (in nanometers-1077.65, 666.94), Normalised Difference Building Index (NDBI)-wavelengths (in nanometers-1788.88, 892.33), Normalised Difference Soil Index (NDSI)-wavelengths (in nanometers-1763.84, 396.47), Clay Mineral Index-wavelengths (in nanometers-2164.54, 1713.75) have been formulated by using the significant wavelengths derived from OIF. Two indices, namely Clay Mineral Index and NDBI have resulted in significantly improving the classification of components/constituents in soil (i.e., Producers’ Accuracy of Arid Soil-92.13% and Non-Arid Soil-93.84%) and man-built (i.e., Producers’ Accuracy of Concrete -83.41% and Asphalt-93.09%) structures. Besides, the research also focuses on improving the classification accuracy of linearly structured man-made object i.e., roads in an urban area by exploring the spatialshape attributes using a very high spatial resolution VIS RGB data. A new approach has been proposed in for calculating the spatial shape features. The results have shown that it is beneficial to study the pixel shape features which can be extended to object level in order to improve the classification accuracy. The classification accuracy achieved through this study for class “road” is 97.3% viii which is the enhanced in comparison to the results achieved so far with other experiments in this research work. Several important conclusions have been drawn from this study regarding improvement in classification accuracy of urban objects and certain new approaches for the same have been recommended. The study also givessuggestions for undertaking future research in certain areas such as for exploring more number of spatial structural features in identifying man made features and use of combination of spectral and spatial features in identification of subclasses or different levels of classification in natural and man-made objects in an urban environment.