Urban Scene Detection by Material Mapping Using Hyper Spectral Imaging

Abstract

Hyperspectral remote sensors collect images in hundreds of narrow adjacent spectral bands. This image data at pixel level is compared with field or laboratory reference spectra in order to recognize and map surface materials present at each pixel. The material mapping is further used to build applications in agriculture, food processing, surface mineralogy, chemical imaging, and surveillance. Techniques have also been proposed in the literature, which utilize material mapping as well as spatial context of materials to recognize a particular target. We propose a novel material-mapping algorithm, which relies on the fact that pixels belonging to the same class but located at different positions in the image exhibit variability in their spectral signatures. This could be due to the difference in terrain, atmosphere and surrounding materials. Therefore, a pixel will better match to the neighboring rather than distant pixels of its own class. The algorithm avoids configuring an SVM and at the same time reduces the complexity of its Nearest Neighbor (NN) style matching scheme. Our algorithm dynamically reduces the training set for each testing pixel. Median matching score of 20 spatially closest members of each class are compared to decide the fate of the testing pixel. Two matching algorithms namely Euclidean distance and Spectral Angle Mapper (SAM) are used. We know that SAM algorithm is robust to multiplicative distortion between test and reference spectra. Our approach is different to, for example, using a Support Vector Machine (SVM). It resembles more to Nearest Neighbor (NN) algorithm. Its complexity is lower than that of NN owing to matching being performed with only limited number of training pixels. In case of SVM, learning takes long especially for long feature vectors. It is generally easier to deal with multiple-class problems with NN than SVM. Several parameters need to be tuned to get good accuracy and generalization from SVM. We use unsupervised learning to help supervised learning by Euclidean and SAM classifiers. The basic idea is that all pixels belonging to a cluster should be classified to the same class. It greatly increases the accuracy of the first stage of our approach. The training data is utilized twice first vi by supervised learning algorithms and then by clustering. If a training pixel is present within a cluster, the whole cluster is classified as belonging to the training pixel class. Our results show that 2nd stage results into comparable accuracy of Euclidean and SAM algorithms. We perform clustering and material classification for the whole images in the datasets. The accuracy, though, is judged only on the ground truth pixels. In most of the cases by target recognition, we mean a target material recognition. Other spatial target may be identified by the cluster analysis of pixels belonging to their material. In some instances, the material of their background may also help, e.g., a bridge is defined as concrete over water. Our first stage uses hard classification of pixels. In the second stage, we have used K-means which provides hard clustering. Fuzzy C Mean (FCM) algorithm provides soft clustering and unmixing techniques provide fuzzy membership to each testing pixel. An interesting dimension would be to use FCM along with unmixing techniques to classify the pixels.