A recommender system-inspired cloud data filling scheme for satellite-based coastal land use classification

Abstract

Filling missing data in cloud-covered areas of satellite imaging is an important task to improve data quantity and quality for enhanced earth observation. Traditional cloud filling studies focused on continuous numerical data such as temperature and cyanobacterial concentration in the open ocean. Cloud data filling issues in coastal imaging have not been studied to the author’s knowledge because of the complex landscape. Inspired by the success of data imputation methods in recommender systems that are designed for online shopping, the present study explored their application to satellite cloud data filling tasks. A numerical experiment was designed and conducted for a Landsat dataset with a range of synthetic cloud covers to examine the performance of different data filling schemes. The recommender system-inspired matrix factorization algorithm called Funk–SVD showed superior performance in computational accuracy and efficiency for the task of recovering landscape types in a complex coastal area than the traditional data filling scheme of DINEOF (Data Interpolating Empirical Orthogonal Functions) and the deep learning method of Datawig. Funk–SVD achieved the best filling accuracy and reached a speed comparable to DINEOF and much faster than deep learning. A theoretical framework was created to analyze the error propagation in DINEOF and found the algorithm needs to be modified to converge to the ground truth. The present study showed that Funk–SVD has great potential to enhance cloud data filling performance and connects the fields of recommender systems and cloud filling to promote the improvement and sharing of useful algorithms.