Urban Dynamic change detection
The Urban Dynamic Processor consists of three modules (see Figure 2):
- Image pre-processing providing completion of missing data due to sensor dropouts and clouds and shadows. The completion and interpolation of missing data module deals with missing data in the Landsat imagery due to sensor dropouts and clouds. Two approaches dealing with missing data include (i) generating complete space-temporal feature vectors data by completion base on clustering and (b) building less descriptive but more restricted composite vectors.
- Change detection model training finds the discriminative hyper-plane providing the optimal discrimination on the training set comprising data examples with ground truth labels. Two classes corresponding to change/no-change labels have been used. Several variations of the feature space have been implemented exploiting spatial and temporal contexts on different scales to incorporate regularization depending on the quality of data.
- Processing providing the detection of changes by classification. Change detection is realized by classifying input data represented by vectors of features for each pixel into the two classes change/no-change. The classification provides for each image pixel the label of the class and a confidence of the classification obtained from the distance of the feature vector from the discriminating hyper-plane in the feature space.
Figure 2: Structure of the processor
The processor is implemented in GNU Octave language/environment and running on VirtualBox virtual machine in 64-bit Linnux CetnOS.
The processor design was based on experimenting with an implementing the state of the art image data modeling (Gaussian mixture models, PCA, clustering and vocabulary models), feature design (temporal and spatial context, robust composite feature construction), and using prior models for classification (Markov processes, Bayesisan chains, kernelized SVM classification). Particular attention has been paid to understanding urban change classes and their correspondence to existing PUMA classes. Independent ground truth labeling using available aerial photographs has been carried out to obtain more reliable ground truth.