Recently, PI from Shanghai Qi Zhi Institute, Professor Zhang Feng from Department of Atmospheric and Marine Sciences of Fudan University and researchers from Nanjing University of Information Science and Technology have established a ground-based cloud detection model TL-DeepLabV3+ by using transfer learning and combining with semantic segmentation network Deeplabv3+. The research results can greatly improve the accuracy and efficiency of ground-based cloud detection tasks, and it is also of great significance for the ground station to realize accurate and efficient automatic cloud identification. Related papers have been published in the journal IEEE Geoscience and Remote Sensing Letters.
The existence, location and height, coverage and development of clouds can predict the weather changes to a certain extent, which plays a vital role in meteorological forecast and disaster early warning and is closely related to social life and national economy. In practical business, the judgment of cloud cover and cloud height is still not divorced from the traditional mode of relying on meteorological personnel's experience judgment, and relying on empirical subjective judgment will inevitably bring many uncertainties. Therefore, it is very important to realize automatic cloud detection of foundation cloud images.
Ground-based cloud images can directly describe low clouds and regional clouds, and specifically describe the external morphological characteristics of clouds. Because of the variety of cloud manifestations and the characteristics of texture similarity, brightness similarity and contour continuity, there are still many shortcomings in traditional cloud detection methods. For example, thin clouds are often ignored in detection, highlighted objects are often mistaken for clouds because of their high reflectivity, and cloud contour detection is inaccurate. With the development of artificial intelligence technology, deep learning has become the first choice in many application fields because of its strong self-learning ability and data analysis ability. In recent years, many researchers have applied the deep learning method, which has made remarkable achievements in the fields of image processing and target detection, to the detection of ground-based cloud images. Professor Zhang Feng's team independently established a ground-based cloud segmentation (GBCS), which contains 1742 ground cloud images and corresponding labels, and is the largest labeled ground cloud image data set at present. They have done a lot of training and comparative evaluation on 12 mainstream deep learning models on this data set. The experimental results show that TL-DeepLabV3+ has achieved very good results in quantitative evaluation, with Pixel Accuracy (PA) reaching 96% and Mean Intersection over Union (MIoU) reaching 91%, all of which are better than other mainstream semantic segmentation models (see Table 1)