Understanding the Role of Receptive Field of Convolutional Neural Network for Cloud Detection in Landsat 8 OLI Imagery

Abstract

Deep semantic segmentation networks perform better in cloud detection of satellite imagery than traditional methods due to their ability to extract high-level features over a large receptive field. However, a large receptive field often leads to loss of spatial details and blurring of boundaries. Therefore, it is crucial to understand the role of the receptive field on the segmentation results, which has rarely been investigated for cloud detection tasks. This study, for the first time, explored the relationship between the receptive field size and the performance of a cloud detection network. Six typical networks commonly used for cloud detection and nine modified UNet variants with different depths, dilated convolutions, and skip connections were evaluated based on the Landsat 8 Biome (L8 Biome) dataset. The theoretical receptive field (TRF) and the effective receptive field (ERF) were introduced to measure the receptive field sizes of different networks. The results revealed a negative correlation between the ERF size and cloud segmentation accuracies for different cloud distributions and a relatively weak negative correlation between the TRF size and segmentation accuracies. Furthermore, ERFs were considerably smaller than the corresponding TRFs for most networks, implying that large-scale contextual information was not learned after training. This result indicates the importance of using networks with a small receptive field for cloud detection of Landsat 8 OLI imagery. Moreover, as the boundary accuracies are significantly lower than the region accuracies, future efforts should be devoted to addressing inaccurate boundary localization rather than exploring the contextual information over a large receptive field.