Abstract: The selection and optimization of features play an important role in the recognition of remote sensing image. Usually, object-based methods cannot make full use of spectral information whereas pixel-based cannot take the advantage of spatial geometry information of remote sensing images. Therefore, a novel urban change detection method is proposed to combine pixel-scale and object-scale features from aerial remote sensing images. First, a feature space was established with spectrum, derivative index, texture, geometry, surface height and convolution neural network layers features. Second, a large number of important features were selected by using LightGBM (Light Gradient Boosting Machine) algorithm. Third, the selected features were used in the random forest classifier to produce the land cover maps from the aerial remote sensing images of Yixing City in 2012 and 2015. Finally, change matrix was used to detect urban change. The results show that the combination of pixel, depth, object features and LightGBM feature selection algorithm has the best recognition effect. Meanwhile, the average accuracy and Kappa coefficient of the proposed method are 88.50% and 0.86, which is 10.50%, 15.00% and 4.00% higher than that based on pixel, deep or object features recognition alone. And the accuracy of urban change detection is 87.50%. Hence, the proposed method is an effective method for urban change detection using aerial remote sensing images.

Key words: combining features of pixel-depth-object, feature from convolutional neural network, LightGBM algorithm, feature selection, urban change detection, aerial remote sensing image