عنوان مقاله [English]
The increasing use of satellite remote sensing data for civilian use has proved to be the most cost-effective means of mapping and monitoring for environmental changes. Satellite remote sensing has played a pivotal role in finding forest cover, vegetation type and land use changes in urban areas. One of the most complete of these methods is classification. The conventional per-pixel image classification techniques have proven ineffective due to disregarding spatial information of the images in digitally classifying urban land-use and land-cover features in high-resolution images. From all classification approaches, texture is believed to be more advantageous for high-resolution images. This is because texture not only utilizes the spectral information but also takes into account the spatial configuration of pixels. The aim of this study is evaluation on the ability of GeoEye-1 data and image texture features and boosted tree classifiers & regression method (BRT) to delineate the urban land cover and urban land use.
Materials and methods
GeoEye-1 images have been employed in the land-use classification. We applied geometric rectification using a road network map. The number of training pixels should at least be equal to ten times the number of variables used in the classification model for a parametric classification approach. However, several studies have shown that non parametric machine learning algorithms require larger number of training data to attain optimal results. To create an exhaustive database with an optimal size for the training and accuracy assessment, 873 sampling points were taken in field surveys using Global Position System (GPS) in the region 3, Tehran City. The ground reference dataset was divided randomly into 7.10 and 3.10 for training and testing, respectively. Then, image texture features including mean and variance of first order, entropy, dissimilarity and homogeneity of second order was processed in ENVI. The Boosted Tree Classifier and Regression (BRT) were used in land use classification. The error matrix of the classification results was formed. The BRT is a combination of statistical and machine learning techniques and an extension of CART, a promising technique used in ecological modeling. Over the past few years, this technique has emerged as one of the most powerful methods for predictive data mining. The BRT combine the strengths of two algorithms: regression trees, models that relate a response to their predictors by recursive binary splits, and boosting, an adaptive method for combining many simple models to give improved predictive performance. It is one of the several techniques that aim to improve the performance of single models by fitting many models and combining them for prediction. The good performance of BRT is depending on regularizing the boosted trees options and stopping tree growing parameters. For boosted tree options, the shrinkage rate as specific weight for single tree and number of boosted trees are two important parameters. Choosing the best shrinkage rate is important to prevent over fitting the predictions. Empirical studies have shown that shrinkage rate of 0.1 or less usually lead to better models. In addition, for small data sets (n=500), the shrinkage rate can be set as 0.005 and for the larger ones (n=5000) it can be set to 0.05. Therefore, regarding the data, the shrinkage rate of 0.05 was used in the present study. The number of boosted trees is effective to produce unbiased results. Thus, to find the optimal tree, initial 300 additive terms trees were set as the number of simple classification trees to be computed in successive boosting steps. For applying the bootstrap training learning, we used 90 percent of training samples. The stopping parameters control the complexity of the individual trees that will be built at each consecutive boosting step. These parameters are including minimum five in child node, which control the smallest permissible number in a child node, for a split to be applied, and maximum fifteen nodes in each tree, which will split.
Result and conclusion
Our results indicated that the overall accuracy and Kappa coefficient for the best compositions of features and main bands were 92% and 90%, respectively. Texture analysis in classification, in fact, the spectral and spatial pattern of pixels were applied simultaneously to obtain better results. As mentioned previously, the texture analysis is capable to increase the accuracy of classification in the especially heterogeneous urban areas.
This can be concluded that the strengths of the GeoEye imagery data and the potentials of the image texture features and BRT method can help the urban planners monitor and interpret complex urban characteristics.