ABSTRACT
Image classification is one of the primary steps for hyperspectral image analysis. The purpose of image classification is to assign a single category label to each pixel in the image. Classification is used for identifying land use land cover of an area. Urban Vegetation Impervious surface Soil (VIS) land cover are hard targets for hyperspectral data analysis as they tend to have signatures similar to each other. Using complete information from the spectrum is helpful.
Unlike land cover, land use in an urban area needs to be inferred as it is not directly observed. Proxy features derived from primary features can be used for land use class effectively. Fractions of VIS within a single pixel or a group of pixels are unique enough to facilitate this. Different economic zones indicate unique VIS fractions vectors.
Spectral Angle Mapper (SAM) is the mainstay of the classifiers for hyperspectral data as it can be easily applied to hyperspectral data from spectrometers and airborne/spaceborne platforms as well. Other classifiers found equally efficient are k-Nearest Neighbour (k-NN), Maximum likelihood, Support Vector Machine (SVM), Random forest, and so on. Modern convolutional neural network (CNN) provides highly accurate classification results on benchmark datasets. The advantage of CNN over conventional classifiers is that it does not require hand-crafted features.
This chapter will discuss the following topics:
o Classification of urban land use and land cover
o Image classification
o Its significance to hyperspectral data
o Supervised and unsupervised machine learning techniques overview
o Challenges in land use land cover classification
o Class codes/models USGS etc.
o Mathematical preliminaries
o Index-based methods
o Method intuition, method
o Different urban indices
o Spectral matching methods
o Structure – general algo
o Similarity measures
o Some of the case studies
o Conventional machine learning (ML, k-NN)
o Simple image features
o Proxy features
o Deep learning (with spectral convolutional focus)
o Spectral, spatial, spectral-spatial features
o Soe common deep learning networks
o Spectral capsules
