Canny Enhanced High-Resolution Neural Network for Satellite Image Based Land Cover Classification and Its Application in Wireless Channel Simulations

Satellite image based land cover classification, which falls under the category of semantic segmentation, is critical for many global and environmental applications. Deep learning has been proven to be excellent in semantic segmentation. However, mainstream neural networks formed by connecting high-to-low convolutions in series are prone to losing image information, which affects the accuracy of semantic segmentation. Besides, it is difficult to distinguish adjacent land cover classes with similar colors using only RGB information presented by satellite images. Striven to maintain high-resolution representations and improve the inter-class distinguishability, a Canny enhanced high-resolution neural network (C-HRNet) is proposed based on Canny edge features and the high-resolution neural network (HRNet) that maintains high-resolution representations throughout the process. Meanwhile, we construct a novel dataset for model evaluation, which provides an automatic dataset construction method, making the dataset more efficient in construction and richer in samples. Extensive experiments are conducted on datasets at different granularities. Quantitative results demonstrate that for large-scale fine-grained scenarios, C-HRNet outperforms state-of-the-art semantic segmentation networks due to the accurate spatial localization ability of Canny edge features. For small-scale coarse-grained scenarios, Canny extracts a large number of edge features that highlight the positional differences between adjacent instances belonging to the same land cover class, which slightly degrades the performance of C-HRNet, but it can still provide reliable land cover classification results. Based on this conclusion, we apply C-HRNet to large-scale wireless channel simulations that are location-sensitive and require fine-grained semantic segmentation, which are proven to be accurate and effective.

Introduction

Land cover is the observed physical cover on the earth's surface. Satellite image based land cover classification is defined as the process of assigning the likelihood of each data pixel in a satellite image to represent certain land cover class, which obviously falls under the category of semantic segmentation, also known as pixel-level image classification [1], [2]. Nowadays, land cover classification has been recognized as a very important research topic in remote sensing as it plays an increasingly critical role in various practical applications, such as global environmental monitoring [3], [4], [5], scenario interpretation [6], and urban expansion [7]. Besides, the study of propagation channel characteristics is the basis for designing and optimizing wireless communication systems [8]. Channel characteristics strongly depend on physical environments, thus land cover classification is also widely used in wireless communication systems [9], [10], [11].

Driven by ever-increasing computing power, machine learning (ML) has achieved great success in basic tasks of computer vision [12]. In particular, the emergence of deep learning (DL) and large available datasets has facilitated tremendous progress in semantic segmentation [13], [14], [15]. To the best of our knowledge, fully convolutional neural network (FCN) is the first neural network to achieve end-to-end classification [16]. Different from convolutional neural networks, the output of FCN corresponds to the input at the pixel level, which specifies the class of each pixel in the input. In [17], an off-the-shelf pretrained FCN is used on synthetic aperture radar images for semantic segmentation. The results show good segmentation results for land use and natural categories, but the segmentation results for buildings are not satisfactory. An encoder-decoder based symmetric U-shaped architecture (U-Net) is proposed in [18], which is an image segmentation technique developed primarily for medical image analysis [19]. DeepLab is one of the most promising techniques for semantic segmentation in DL, with several advanced versions after several improvements [20], [21], [22], [23]. DeepLab consists of encoder and decoder. The former utilizes convolutional layers to extract low-resolution representations, and the latter utilizes upsampling to restore the resolution of representations to that of input images for semantic segmentation.