28 Jun 2024
Research Update

Masters students uncover solutions to Cloud Cover in Hyperspectral Imagery

28 Jun 2024

SmartSat was proud to be the Project Sponsor for a post-graduate student project through the University of South Australia looking into Cloud Cover Detection from Hyper-Spectral Imagery.

Cloud cover segmentation enables a more comprehensive view of Earth’s surface features, facilitating greater accuracy in analysis, interpretation, and decision-making. By removing the influence of cloud cover, SmartSat can extract valuable insights from satellite imagery. Stakeholders who depend on this information are then in a better position to make more informed policy and business decisions.

Six members from the University of South Australia had been engaged to deliver a solution: Ainu Joshy, Esha Trivedi, Ram Sai Reddy Atla, Robert Flacco, Serah Johnson, Shubham Dogra. The project objective was to determine the best machine learning algorithm for cloud cover detection in hyperspectral satellite images, according to various evaluation criteria subject to computational complexity and efficiency constraints and deliver this to SmartSat together with its associated documentation.

By analysing the relationship between reflectance values of satellite images and class categories of their pixels, (cloud v non-cloud), the team aimed to develop a series of competing machine learning algorithms for cloud cover detection. Their project evaluated five alternative machine learning algorithms for cloud cover detection in hyper-spectral satellite images. Two groups of dataset images were tested by the machine learning algorithms which include four trained images and four untrained images. The machine learning algorithms utilise ten-fold cross-validation repeated three times to find the optimal hyperparameter setting for performance maximisation.

Performance results indicate that Decision Trees achieved the best all-round performance metrics for cloud cover in hyper-spectral images, as well as displayed the least volatility among the numerous performance metrics including accuracy, precision, recall, F1, and Kappa.

Feature image: University of South Australia students (L-R) Shubham Dogra, Ainu Joshy, Robert Flacco, Esha Trivedi, Serah Johnson and Ram Sai Reddy Atla with SmartSat Chief Operations Officer, Andrew Beveridge (second from left).