Supervised and Adaptive Feature Weighting for Object-Based Classification on Satellite Images

Object-based image analysis (OBIA) technique has been representing an evolving paradigm of remote sensing application, along with more high-resolution satellite images available. However, too many derived features from segmented objects also present a new challenge to OBIA applications. In this paper, we present a supervised and adaptive method for ranking and weighting features for object-based classification. The core of this method is the feature weight maps for each land type resulted from prior thematic maps and their corresponding satellite images of study areas. Specifically, first, satellite images to be classified are segmented using an adaptive multiscale algorithm, and the multiple (spectral, shape, and texture) features of segmented objects are calculated. Second, we extract distance maps and feature weight vectors for each land type from the prior thematic maps and corresponding satellite images, to generate feature weight maps. Third, a feature-weighted classifier with the feature weight maps, is applied on the segmented objects to generate classification maps. Finally, the classification result is evaluated. This approach is applied on a Sentinel-2 multispectral satellite image and a Google Map image to produce objected-based classification maps, compared with the traditional feature selection algorithms. The experimental results illustrate that the proposed method is practically efficient to select important features and improve classification performance.