Spectral-textural Image Classification in a Semiarid Environment

Image classification can benefit from incorporating texture by enabling an increased number of classes and improving thematic accuracy. Incorporating texture also involves special attention in a number of aspects that range from the texture source to the evaluation of accuracy through pre-processing, training strategy and choosing a texture extraction paradigm and a classifier. Without special care in these aspects, classification results can be very unpredictable, especially when mixing spectral and textural features in the classification. This is mainly due to the spatial dependency of texture features. The present article aims at analyzing these aspects (six in all) through a review of the concepts involved and a demonstration with two sample image data sets in a complex semiarid environment in Brazil. The data sets were formed with texture features from a SPOT-5 panchromatic image and spectral features from LANDSAT 7 ETM+ data. Results suggest that useful texture features can be extracted from SPOT-5 panchromatic data and that a mixed classification scheme is generally better than either approaches (spectral or textural). They also suggest that a non parametric classifier (Fisher linear discriminant) performs better for sets incorporating spectral and textural features and is less affected by edges and borders.