Pixel-Wise Classification Method for High Resolution Remote Sensing Imagery Using Deep Neural Networks

Abstract

Considering the classification of high spatial resolution remote sensing imagery, this paper presents a novel classification method for such imagery using deep neural networks. Deep learning methods, such as a fully convolutional network (FCN) model, achieve state-of-the-art performance in natural image semantic segmentation when provided with large-scale datasets and respective labels. To use data efficiently in the training stage, we first pre-segment training images and their labels into small patches as supplements of training data using graph-based segmentation and the selective search method. Subsequently, FCN with atrous convolution is used to perform pixel-wise classification. In the testing stage, post-processing with fully connected conditional random fields (CRFs) is used to refine results. Extensive experiments based on the Vaihingen dataset demonstrate that our method performs better than the reference state-of-the-art networks when applied to high-resolution remote sensing imagery classification. accuracy in certain situations. Considering the complexity of land cover that contains vegetation, water, soil and other physical land features, including those created solely by human activities, it is still challenging for object-oriented approaches to improve classification accuracy.Traditionally, classification of remote sensing imagery focused on two steps, i.e., feature extraction methods and supervised learning algorithms. Various methods have been proposed to extract features using hand-crafted descriptors, such as Scale-Invariant Feature Transform (SIFT) [11] and Histogram of Oriented Gradients (HOG) [12]. However, the two steps mentioned above are typically viewed as distinct approaches. Convolutional neural networks (CNNs) combine the two steps into one network that learns contextual features at different scales and computes the score of each class at the end of the network. Another benefit of CNNs is that they can be trained to optimize all weights in the network using an efficient weight update technique such as stochastic gradient descent (SGD) [13] end-to-end. Although CNNs have achieved remarkable results in image categorization, they do not consider pixel-wise semantic classification. In image categorization, the input of a network is an image, while the goal is to predict the correct category of the entire image. FCN [14]-based models have been successfully applied to natural image semantic segmentation and are currently the state-of-the-art for PASCAL VOC datasets [15]. However, FCN models do not take into account relations between pixels and ignore the spatial regularity in the usual pixel-wise classification method. In addition, the vast majority of research focuses on optimization of the network structure without considering efficient use of labeled data that is very important to deep neural networks.To overcome the above challenges, in this paper, we introduce a novel pipeline for pixel-wise classification of high-resolution remote sensing imagery. Our method consists of three parts: ...

DV3+HED+: a DCNN-based framework to monitor temporary works and ESAs in railway construction project using VHR satellite images

“…Marmanis [32] extracted scale-dependent class boundaries before each pooling level, with the class boundaries fused into the final multi-scale boundary prediction. Guo [33] extracted bounding boxes of potential ground objects which augmented the training dataset before training the DCNNs. Tian [34] presented DFCNet(Dense Fusion Classmate Network) which was jointly trained with auxiliary road dataset properly compensates the lack of mid-level information.…”

A Dual-Path and Lightweight Convolutional Neural Network for High-Resolution Aerial Image Segmentation

“…This method achieved high overall accuracy as well as good performance for small objects segmentation. Guo et al [19] exploited FCN with atrous convolution to perform semantic segmentation for high-resolution remote sensing images. They used graph-based segmentation and selective search method to augment the training data and conditional random fields(CRF) to refine the segmentation results.…”

Edge Prior Multilayer Segmentation Network Based on Bayesian Framework

“…FCN modifies CNN to obtain the classification results of each pixel and implement semantic segmentation. In general, CRF usually serves as a postprocessing method of FCN [17,18] to improve the segmentation results and DeepLab is a typical case. DeepLab performs semantic segmentation with atrous convolution, deep convolutional nets, and fully connected CRFs.…”