Abstract
A new set of temporal features derived from codispersion and cross-semivariogram geostatistical functions is proposed, described, extracted, and evaluated for object-based land-use/land-cover change detection using high resolution images. Five features were extracted from the codispersion function and another six from the cross-semivariogram. The set of features describes the temporal behaviour of the internal structure of the image objects defined in a cadastral database. The set of extracted features was combined with spectral information and a feature selection study was performed using forward stepwise discriminant analysis, principal component analysis, as well as correlation and feature interpretation analysis. The temporal feature set was validated using high resolution aerial images from an agricultural area located in south-east Spain, in order to solve a tree crop change detection problem. Direct classification using decision tree classifier was used as change detection method. Different classifications were performed comparing various feature group combinations in order to obtain the most suitable features for this study. Results showed that the new sets of cross-semivariogram and codispersion features provided high global accuracy classification results (83.55% and 85.71% respectively), showing high potential for detecting changes related to the internal structure of agricultural tree crop parcels. A significant increase in accuracy value was obtained when combining features from both groups with spectral information (94.59%).
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have