Flowchart for the artificial neural network classification. 

Contexts in source publication

Context 1

... QuickBird II satellite images with 2.44 m (multitespectral image) and 61 cm (panchromatic image) spatial resolutions were obtained (overpass: February, 2007). These images were acquired in standard correction level and were made available by Professor Dr. Amilton Amorim from the Department of Cartography of the FCT/UNESP, through research project funded by the São Paulo Research Foundation (FAPESP). The images were georeferenced to the UTM projection system and SAD69 datum considering a linear transformation and interpolation by the nearest neighborhood. A set of 20 control points identified in the vector database were used to georeference the panchromatic image, achieving an accuracy of 3.14 pixels. Taking the corrected panchromatic image as a reference, another 20 control points were used to register the multispectral image, resulting in an error of 0.42 pixels. Panchromatic and multispectral images were then merged using the Gram-Schmidt technique (Laben and Brower, 2000). Then, it was generated an occurrence texture image using an entropy filter with window size of 9 pixels by 9 pixels. This procedure allows the combination of spatial and spectral information in the classification process, minimizing the spectral similarity of some targets. Pre- liminary tests conducted with conventional classifiers indicated, for example, confusion between asphalt and asbesto-cement and between soil and red ceramics. The four fused Quickbird bands and texture image were used to perform an artificial neural network (ANN) classification. The steps required for the thematic information extraction are illustrated in Figure 3. Training and validation samples were obtained for each land cover class. The architecture specified was a multilayer feedforward network, which requires the definition of an input layer, one or more hidden layers and output layer. Simulations were carried out with one or two hidden layers. The performance of the trained network was verified with backpropagation learning algorithm. The number of nodes in the input layer was defined by the size of input data, that is, the four fused bands and the texture image with five nodes. The supervised training process using backpropagation involves a prior selection of all samples and indication of the number of training and validation pixels to be considered for each class. Furthermore, it was necessary to define in advance the training parameters, which includes the learning rate, the momentum factor and the stopping criterium. The last one controls the end of the process and includes three aspects: the value of the acceptable mean square error (MSE), the number of iterations and the accuracy rate. Several simulations were performed varying the number of iterations and analyzing the training statistics, which included: values of the test and validation errors, number of iterations and accuracy rate computed from the test and validation pixels collected for the classes. After training the ANN, the classification was performed generating conventional (hard) and soft outputs of the thematic classes, which express the membership degrees (activation levels) of the pixels in relation to each class. In order to minimize the number of isolated pixels within the classified regions and pixels located in the edge regions (transition between classes), where a large amount of misclassified pixels can occur, the mode filter was applied on the classified image, considering a 5 × 5 window and setting the final classification result. To assess the reliability of the urban scale mapping, statistical indices (overall and by class) and their corresponding uncertainty measurements were calculated. The reference data were based on three-day field survey. The random sample unit was a cluster of 3 pixels by 3 pixels. The number of sample elements per class was calculated using Equations 1 and 2. The analysis of mapping uncertainty was based on the soft images that express the activation levels of output for each pixel for each class, which are ...

Context 2

... four fused Quickbird bands and texture image were used to perform an artificial neural network (ANN) classification. The steps required for the thematic information extraction are illustrated in Figure 3. ...