关键词: 深度学习, 地理国情普查, 地表覆盖分类, 质量评价, AlexNet

Abstract:

As one of the important outcomes of the National Geographic Census of China, the land cover classification reveals the information of both natural and artificial coverage elements, including vegetation, soil, glaciers, rivers, lakes, marsh wetlands and various artificial structures. Obviously, it mainly focuses on profiling the natural characters of the land surface with temporal and distribution attributes, which has an obviously different classification system from other scene classification applications. In recent years, more and more high-resolution remote sensing platforms become available, it is possible to update and evaluate land cover classification quickly with the advantage of huge volume of data and more frequent data updates. Meanwhile, in practice we face with more and more challenges of the huge data. In this paper, we propose a novel approach for evaluating the land cover classification results by combining the object-oriented method with the Deep Convolutional Neural Network (D-CNN) model. With deeper structure and wilder receptive field, the deep neural network has the capability of abstract description from low-level features, and the deep learning has become one of the latest development trends in the artificial neural network field. The deep learning shows a completely different possibility in many fields, and it has been applied to the speech recognition, image recognition, information retrieval and so on. The newly-developed method of image recognition based on deep leaning has been preliminarily verified in the scene classification field. Traditionally, the land cover classification method is established on the pixel-based classifying. The latest improved method of the object-oriented classification frame has been proposed, but this new frame is hard to be achieved because of the lack of supports from efficient methods and algorithms. Nowadays, the deep neural network provides us an effective tool to achieve the object-oriented classification by clipping image spots from original images and inputting the clipped image spots to D-CNN. The D-CNN model can convolute and pool the image spots to realize the object-oriented classification of the land cover. By the combination of the object-oriented classification with the deep learning, the proposed method can extract more and better abstract features than the pixel-based approach, while the pixel-based method requires more manual interventions. When applying the deep learning method to land cover classification recognition, the prepared image spots as appropriate inputs will be automatically scored to its belonging classes. Thus, the score represent the degree of membership of the image spot matching to the corresponding class. By fine-tuning the D-CNN, we can obtain a new approach of judging the quality of the samples, in order to assure the reliability of the proposed approach. The fine-tuned D-CNN is required to be sufficiently robust, and we verify its robustness in the following experiment by employing the AlexNet. The experimental results show that the image spots of arable land and building can be recognized with the membership degree of 99.95% and 99.41%, but those of woodland and water area can be recognized only with the membership degree of 62.73% and 43.59%. Obviously, the proposed model can achieve the promising reliability that is related to the qualified and sufficient data set of the image spots which is used for fine-tuning of the net. The reason for poor robustness of the fine-tuned AlexNet in classifying the woodland and water area may be the insufficient size of data-set of these two classes. It shows that a fine-tuned deep convolutional neural network as a new model can be utilized in evaluating the land cover classification with high reliability.

Key words: deep learning, national geographic census, land cover classification, quality evaluation, AlexNet