Multiclass Cloud Detection and Segmentation in Satellite Hyperspectral Imagery

Date of Award

5-9-2026

Degree Name

M.S. in Electrical Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Vijayan Asari

Abstract

At any given time 67% of the Earth’s surface is covered by clouds. This presents a challenge for satellite imagery data analysis as clouds, particularly dense clouds, are a common obstruction of the ground. Ideally, in scenes with high levels of clouds, the clouds could be removed or masked out to avoid interference with work focused on surface data. In this work we develop and evaluate several methods for cloud detection and segmentation in hyperspectral imagery (HSI). Hyperspectral imagery contains hundreds of narrow continuous bands which introduces unique advantages and challenges. This work evaluates the performance of four encoder-decoder architectures: U-Net, ResUNet, RU-Net, R2U-Net for two-class cloud detection in hyperspectral imagery. The objective is to differentiate dense, optically thick clouds, from thin semi-transparent clouds. These architectures were selected for their prowess in segmentation and introduction of residual shortcuts and/or recurrence to aid in the segmentation of the thin cloud boundaries. Different band selection criteria, steps per epoch and if applicable recurrent iterations are tested to fnd the best combination. Experiments are conducted on data cubes from the Earth Surface Mineral Dust Source Investigation (EMIT) sensor. EMIT hyperspectral data cubes contain 285 spectral bands and a ground sample distance of 60 meters. The truth is manually labeled using the MATLAB Image Labeler App with two classes: thin clouds and dense clouds. The best performing model is found to be the RU-Net trained on every 5th band and 600 steps per epoch with an average testing IoU of 0.8132 for dense clouds and 0.6430 for thin clouds. When one class results were compared to the EMIT cloud masks the RU-Net has an average +153.43% improvement in IoU, +94.54% in F1-score and +25.44% in accuracy. When compared to the dense cloud masks there was an average +27.00% increase in IoU and a +16.74% improvement in F1-score. The results demonstrate that this is a viable method for cloud segmentation in HSI showing signifcant improvement over baseline methods.

Keywords

Electrical Engineering

Comments

OCLC No. 1591818283

Rights Statement

Copyright 2026, author.

Link to Full Text

Share

COinS
 
 
 

Links