Cloud detection based on deep neural network for HY-1C COCTS

Presented at the GHRSST XXIII international science team meeting, 27 June-1 July 2022, online and in-person (Barcelona). #GHRSST23

Short abstract

The Haiyang-1C (HY-1C) satellite is the first operational ocean color satellite of the Chinese HY-1 series satellites. The Chinese Ocean Color and Temperature Scanner (COCTS) onboard the HY-1C satellite has 10 channels for ocean color and sea surface temperature (SST) observations. Cloud detection is one of the key pre-processing steps of SST retrieval. Deep learning algorithm can combine spectral information and spatial information and has strong ability of feature extraction. The U-Net is one of useful convolutional networks for image segmentation, consisting of the encoder and the decoder. We use the deep learning model U-Net to identify the cloud over the ocean in HY-1C COCTS images. The HY-1C COCTS cloud detection dataset is composed of HY-1C COCTS L1B global area coverage data in August 2019. The dataset is randomly divided into training set, eval set and test set with the ratio of 7:2:1. The ground truth of dataset using to train the U-Net model is constructed by Bayesian cloud detection method and manual mask. The overall accuracy on test dataset of the deep learning method is 0.95. The SST retrieval based on Optimal Estimation (OE) algorithm for clear pixels detected by U-Net is conducted. The COCTS OE SSTs are compared with iQuam in situ SST. The bias and standard deviation of the COCTS minus in situ SST difference are -0.1 K and 0.53 K, respectively. The ratio of matchups with SST difference lower than -1.67 K is 1.11%, indicating that the missed cloud detection is not obvious. In addition, the brightness temperature images after cloud detection show that the performance of cloud detection over the ocean front works well. In general, the cloud detection based on deep learning algorithm performs well for HY-1C COCTS.