| Abstrakt: |
Despite using multiple input datasets for effective crop classification, it is important to select an appropriate method that efficiently integrates these multiple datasets to produce accurate classification results. In this paper, we present an endorsement theory-based crop classification approach that considers the qualitative information, in terms of prediction probabilities, from different input datasets and integrates them efficiently to produce final classification results. Three different input datasets are used in this study: 1) spectral; 2) texture; and 3) indices from multitemporal (spring, early summer) WorldView-2 multispectral imagery. A multilayer perceptron classifier is trained with the multitemporal datasets separately using a backpropagation learning algorithm, and prediction probabilities are produced for each pixel as evidence against each crop class. An integration rule based on endorsement theory is applied to these multiple evidence by considering their individual contribution, and the most probable class of a pixel is identified. Integration of the three multidate datasets by the proposed method is found to produce higher overall classification accuracy (91.2%) when compared to conventional winner-takes-all approach (89%). In order to determine which individual dataset is more useful for crop discrimination, the dataset's performance is compared using evidence and contributions produced in the proposed integration method for four selected crops, for both single- and multidate. The results of this analyses showed that seasonal textures information outperformed both spectral and indices. To verify this finding, results of individual dataset classification are examined. The highest overall classification accuracy of 88.8% is achieved by the use of multidate texture, where multidate spectral and indices resulted in 86.3% and 84.4%, respectively. [ABSTRACT FROM PUBLISHER] |
