This paper presents a case study on the use of features derived from remote sensing data for mapping the highly fragmented semideciduous Atlantic forest in Brazil. Innovative aspects of this research include the evaluation of different feature sets in order to improve land cover mapping. The feature sets were defined based on expert knowledge and on data mining techniques to be input to traditional and machine learning algorithms for pattern recognition, viz. maximum likelihood, univariate decision trees, multivariate decision trees, and neural networks. The results showed that the maximum likelihood classification using temporal texture descriptors as extracted with wavelet transforms was most accurate to classify the semideciduous Atlantic forest. In this study, a special accuracy measure was used: the so-called class mapping accuracy. Maximum likelihood performed relatively well, with forest mapping accuracies ranging from 34.5 to 51.3%. In contrast, accuracies for neural networks ranged from 19.0 to 45.2%. Classification confusion occurred mainly with coffee and eucalyptus plantations. Univariate trees provided the most robust results for different feature sets, with accuracies ranging from 39.6 to 46.7%. Temporal information of vegetation indices was more important than image texture, terrain topography and raw spectral information for discriminating semideciduous Atlantic forest.
|Number of pages||14|
|Journal||International Journal of Applied Earth Observation and Geoinformation|
|Publication status||Published - Sep 2004|
- artificial intelligence
- feature sets
- forest classification