Mapping and monitoring forest remnants: a multiscale analysis of spatio-temporal data

KEYWORDS : Landsat, time series, machine learning, semideciduous Atlantic forest, Brazil, wavelet transforms, classification, change detection Forests play a major role in important global matters such as carbon cycle, climate change, and biodiversity. Besides, forests also influence soil and water dynamics with major consequences for ecological relations and decision-making. One basic requirement to quantify and model these processes is the availability of accurate maps of forest cover. Data acquisition and analysis at appropriate scales is the keystone to achieve the mapping accuracy needed for development and reliable use of ecological models. The current and upcoming production of high-resolution data sets plus the ever-increasing time series that have been collected since the seventieth must be effectively explored. Missing values and distortions further complicate the analysis of this data set. Thus, integration and proper analysis is of utmost importance for environmental research. New conceptual models in environmental sciences, like the perception of multiple scales, require the development of effective implementation techniques. This thesis presents new methodologies to map and monitor forests on large, highly fragmented areas with complex land use patterns. The use of temporal information is extensively explored to distinguish natural forests from other land cover types that are spectrally similar. In chapter 4, novel schemes based on multiscale wavelet analysis are introduced, which enabled an effective preprocessing of long time series of Landsat data and improved its applicability on environmental assessment. In chapter 5, the produced time series as well as other information on spectral and spatial characteristics were used to classify forested areas in an experiment relating a number of combinations of attribute features. 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 and multivariate decision trees, and neural networks. The results showed that maximum likelihood classification using temporal texture descriptors as extracted with wavelet transforms was most accurate to classify the semideciduous Atlantic forest in the study area. In chapter 6, a multiscale approach to digital change detection was developed to deal with multisensor and noisy remotely sensed images. Changes were extracted according to size classes minimising the effects of geometric and radiometric misregistration. Finally, in chapter 7, an automated procedure for GIS updating based on feature extraction, segmentation and classification was developed to monitor the remnants of semideciduos Atlantic forest. The procedure showed significant improvements over post classification comparison and direct multidate classification based on artificial neural networks.