Transferring Deep Convolutional Neural Networks for the Scene Classification of High-Resolution Remote Sensing Imagery

Abstract

Abstract:Learning efficient image representations is at the core of the scene classification task of remote sensing imagery. The existing methods for solving the scene classification task, based on either feature coding approaches with low-level hand-engineered features or unsupervised feature learning, can only generate mid-level image features with limited representative ability, which essentially prevents them from achieving better performance. Recently, the deep convolutional neural networks (CNNs), which are hierarchical architectures trained on large-scale datasets, have shown astounding performance in object recognition and detection. However, it is still not clear how to use these deep convolutional neural networks for high-resolution remote sensing (HRRS) scene classification. In this paper, we investigate how to transfer features from these successfully pre-trained CNNs for HRRS scene classification. We propose two scenarios for generating image features via extracting CNN features from different layers. In the first scenario, the activation vectors extracted from fully-connected layers are regarded as the final image features; in the second scenario, we extract dense features from the last convolutional layer at multiple scales and then encode the dense features into global image features through commonly used feature coding approaches. Extensive experiments on two public scene classification datasets demonstrate that the image features obtained by the two proposed scenarios, even with a simple linear classifier, can result in remarkable performance and improve the state-of-the-art by a significant margin. The results reveal that the features Remote Sens. 2015, 7 14681 from pre-trained CNNs generalize well to HRRS datasets and are more expressive than the low-and mid-level features. Moreover, we tentatively combine features extracted from different CNN models for better performance.

Very Deep Convolutional Neural Networks for Complex Land Cover Mapping Using Multispectral Remote Sensing Imagery

“…This increases the depth of the network and contributes to learning more complex features. The impressive results of VGG revealed that the network depth is an important factor in obtaining high classification accuracy [32]. …”

“…Unfortunately, wetlands have undergone variations due to natural processes, such as changes in temperature and precipitation caused by climate change, coastal plain subsidence and number of training samples constrains the efficiency of this technique due to the overfitting problem. The other two strategies are more useful when a limited amount of training samples is available [26,32].…”

Multiscale Geoscene Segmentation for Extracting Urban Functional Zones from VHR Satellite Images

“…Then, middle-level features, including object semantics [4,8], visual elements [7], and bag-of-visual-word (BOVW) representations [13], are more effective than low-level features in representing functional zones [7], but they ignore spatial and contextual information of objects, leading to inaccurate recognition results. To resolve this issue, Hu et al (2015) extracted high-level features using convolutional neural network (CNN) [10], which could measure contextual information and were more robust than visual features in recognizing functional zones [14,16]. Zhang et al (2017) had a different opinion on the relevance of deep-learning features and stated that these features rarely had geographic meaning and were weak for the purpose of interpretability [4].…”

“…Finally, functional zones can be labeled with categories based on their features. Previous efforts at functional-zone analysis focus mainly on feature representations [9][10][11][12][13][14] and classification methods [4,15,16], but ignore zone segmentation. This is unfortunate because zone segmentation is an essential precursor to the other two steps of functional-zone analysis and is hence fundamental to the entire undertaking.…”

Four Decades of Estuarine Wetland Changes in the Yellow River Delta Based on Landsat Observations Between 1973 and 2013

“…The technical errors were mainly caused by the characteristics of the wetlands, such as changes in different season, they have complex spectra, they are heterogeneous, and the same land cover types have multiple spectrums. To improve the classification accuracy in the future, more research can be done on the following aspects: (1) To discover more effective features, not just in spectrum, new technologic methods maybe good alternative choices (e.g., synthetic aperture radar, lidar, and geospatial modeling); (2) to enhance the intensity of machine learning; taking into account the all possible situations via the new learning structures (e.g., deep convolutional artificial neural network (ANN) and deep learning) [104][105][106][107][108][109][110][111][112]. The deep convolutional neural network algorithm, in particular, has better learning and generalization performance for multiple variables and large datasets.…”