Automatic Features Extraction From Time Series Of Passive Microwave Images For Snowmelt Detection Using Deep-Learning – A Bidirectional Long-Short Term Memory Autoencoder (Bi-Lstm-Ae) Approach.

• Main Author: Massamba, Bienvenu Sedin
• Format: Text
• Language: unknown
• Published: LSU Scholarly Repository 2020
• Subjects: Micowave remote sensing; Deep-leanring; time-series; machine learning; Geographic Information Sciences; Geography; Human Geography; Physical and Environmental Geography; Remote Sensing; Spatial Science; Antarc*; Antarctic; Antarctica; Ice Sheet
• Online Access: https://repository.lsu.edu/gradschool_theses/5129; https://doi.org/10.31390/gradschool_theses.5129; https://repository.lsu.edu/context/gradschool_theses/article/6179/viewcontent/Massamba_thesis.pdf

The Antarctic surface snowmelt is prone to the polar climate and is common in its coastal regions. With about 90 percent of the planet's glaciers, if all of the Antarctica glaciers melted, sea levels will rise about 58 meters around the planet. The development of an effective automated ice-sheet snowmelt monitoring system is therefore crucial. Microwave remote sensing instruments, on the one hand, are very sensitive to snowmelt and can see day and night through clouds, allowing us to distinguish melting from dry snow and to better understand when, where, and for how long melting has taken place. On the other hand, deep-learning (DL) algorithms, which can learn from linear and non-linear data in a hierarchical way robust representations and discriminative features, have recently become a hotspot in the field of machine learning and have been implemented with success in the geospatial and remote sensing field. This study demonstrates that deep learning, particularly long-short memory autoencoder architecture (LSTM-AE) is capable of fully exploiting archives of passive microwave time series data. In this thesis, An LSTM-AE algorithm was used to reduce and capture essential relationships between attributes stored as brightness temperature within pixel time series and k-means clustering is applied to cluster the leaned representations. The final output map highlights the melt extent in Antarctica.