DETECTION OF CLOUD SHADOWS USING DEEP CNN UTILISING SPATIAL AND SPECTRAL FEATURES OF LANDSAT IMAGERY

Abstract

Background: Cloud and cloud-shadow contamination degrades the usability of Landsat satellite imagery, and existing masking approaches such as fuzzy-c and f-masking leave room for more accurate, pixel-level detection.
Objective: This work focuses on detecting cloud shadows using a deep Convolutional Neural Network (CNN) that utilises the spatial and spectral features of Landsat imagery.
Methods: A deep CNN with approximately 13 hidden layers is built using Python libraries and trained on Landsat images processed through preprocessing, training, CNN, up-sampling, and pooling modules, with a Fourier transform applied to the feature array.
Results: The deep CNN algorithm combines the whole input image to obtain multilevel features, and was evaluated against ground-truth masks across a range of prediction scenarios.
Conclusion: The Deep CNN algorithm performs better image processing and semantic segmentation than existing fuzzy-c and f-masking approaches.

Keywords: Convolutional Neural Network (CNN), Cloud Detection, Semantic Segmentation, Satellite Imager, Landsat, Fuzzy-C.

HIGHLIGHTS:
1.    An improved approach using Deep CNN performs better image processing and semantic segmentation than existing fuzzy-c and f-masking methods.
2.    A five-module pipeline - preprocessing, training, CNN, up-sampling, and pooling - detects clouds and their shadows in Landsat imagery.
3.    The model segments each pixel into cloud, dark cloud shadow, and light cloud shadow, evaluated against ground-truth masks across multiple prediction scenarios.