P_Segnet and NP_Segnet: New Neural Network Architectures for Cloud Recognition of Remote Sensing Images

Junyu Lu, Yuzhu Wang, Yueqin Zhu, Xiaohui Ji, Tingyan Xing, Wei Li, Albert Y. Zomaya

Introduction

Remote sensing is a critical tool for global observation, supporting applications from weather forecasting to agricultural planning. However, persistent cloud cover—affecting over 50% of the Earth's surface—presents a significant challenge for data accuracy. Traditional cloud detection methods rely on manual thresholding of brightness or temperature, which are labor-intensive and often fail to distinguish clouds from other high-reflectance surfaces like snow or bright infrastructure. This research introduces two new neural network architectures, P_Segnet and NP_Segnet, designed to automate and improve the precision of cloud-snow discrimination in satellite imagery.

Methodology and System Architecture

The proposed models build upon the classic SegNet encoder-decoder framework. The primary innovation is the integration of parallel convolutional layers inspired by GoogLeNet’s Inception modules. By using a multi-scale filter bank (3×3 and 5×5 filters) in the initial layers, the networks can fuse features from different receptive fields, allowing them to capture both fine edges and broader contextual information simultaneously.

System workflow for cloud recognition neural network training and deployment.

Two specific variants were tested:

  • P_Segnet (Parallel SegNet): Adds parallel structures only to the encoder portion of the network.
  • NP_Segnet (New Parallel SegNet): Implements a symmetric architecture by adding parallel structures to both the encoder and the decoder.

To ensure computational efficiency despite the increased network width, the researchers utilized 1×1 convolutions for dimension reduction and Batch Normalization to stabilize and speed up the training process.

Dataset and Labeling

The study utilized a high-quality custom dataset consisting of 670 training images and 342 testing images in RGB format. These were sourced from the China Geological Survey and supplemented by web crawling. A critical phase of the research involved the manual labeling of these images into three distinct classes: Cloud, Snow, and Background.

The sequential process for generating image labels and preparing the dataset.

To address the "small data" challenge, the authors employed a pre-training strategy on a representative subset of 50 images with a high learning rate before training on the full dataset. A specialized evaluation subset of 133 "cloud-snow mixed" images was used to specifically test the models' ability to handle the most difficult discrimination tasks.

Key Results and Insights

The experimental results highlight the advantages of parallel, symmetric architectures for semantic segmentation in remote sensing:

  • Cloud Accuracy: NP_Segnet achieved a cloud recognition accuracy of 91.2%, outperforming the baseline SegNet (88.6%) and significantly surpassing standard Fully Convolutional Networks (FCNNs) at 80.4%.
  • False Positive Reduction: While SegNet maintained a high raw accuracy for snow, NP_Segnet was superior at reducing false positives, meaning it was far less likely to misidentify bright snow as cloud cover.
  • Boundary Precision: Qualitatively, the parallel architectures produced much sharper and smoother cloud boundaries compared to the blurry, ambiguous results produced by FCNNs.

The findings suggest that the symmetric encoder-decoder structure of NP_Segnet is vital for maintaining robust discrimination when dealing with highly similar spectral signatures.

Limitations and Future Work

Despite the performance gains, the authors noted several areas for improvement. The current models are memory-intensive due to their increased width and require fixed input image sizes, which can lead to detail loss during resizing. Future research will focus on expanding the dataset to include more diverse cloud formations (such as cirrus and cumulus) and optimizing the architecture for real-time processing on standard hardware.

Summary updated

Abstract

In recent years, remote sensing images have played an important role in environmental monitoring, weather forecasting, and agricultural planning. However, remote sensing images often contain a large number of cloud layers. These clouds can cover a large amount of surface information. Therefore, an increasing number of cloud recognition methods have been proposed to reduce the impact of cloud cover. There are many difficulties in traditional cloud recognition methods. For example, the threshold method based on spectral features improves the accuracy of cloud detection, but it often leads to omission or misjudgment in cloud detection and depends on prior knowledge. To improve the accuracy and efficiency of cloud recognition, we use deep learning to address cloud recognition problems in remote sensing imagery. We propose a series of methods from the acquisition and production of training datasets to neural network training and cloud recognition applications. This paper describes a realization of cloud recognition of remote sensing imagery based on SegNet architecture. We have proposed two architectures named P_SegNet and NP_SegNet, which are modified from SegNet. Some parallel structures were also employed into the SegNet architecture to improve the accuracy of cloud recognition. Due to these changes, this paper also discusses the impact of the symmetry network structure on the final accuracy. Our proposed method was compared with the well-known fully convolutional neural network (FCNN) approach. The results have demonstrated the feasibility and practicality of using deep learning approach for cloud recognition in remote sensing images.
BibTeX
@article{lu2019p,
  author = {Junyu Lu and Yuzhu Wang and Yueqin Zhu and Xiaohui Ji and Tingyan Xing and Wei Li and Albert Y. Zomaya},
  title = {{P\_Segnet and NP\_Segnet: New Neural Network Architectures for Cloud Recognition of Remote Sensing Images}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2925565},
}
Figure