A carregar...
Publicação
Extracting surface water bodies from sentinel-2 imagery using convolutional neural networks
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 12.07 MB | Adobe PDF |
Autores
Resumo(s)
Water is an integral part of eco-system with significant role in human life.
It is immensely mobilized natural resource and hence it should be monitored
continuously. Water features extracted from satellite images can be utilized for
urban planning, disaster management, geospatial dataset update and similar
other applications. In this research, surface water features from Sentinel-2
(S2) images were extracted using state-of-the-art approaches of deep learning.
Performance of three proposed networks from different research were assessed
along with baseline model. In addition, two existing but novel architects
of Convolutional Neural Network (CNN) namely; Densely Convolutional Network
(DenseNet) and Residual Attention Network (AttResNet) were also implemented
to make comparative study of all the networks. Then dense blocks,
transition blocks, attention block and residual block were integrated to propose
a novel network for water bodies extraction. Talking about existing networks,
our experiments suggested that DenseNet was the best network among them
with highest test accuracy and recall values for water and non water across all
the experimented patch sizes. DenseNet achieved the test accuracy of 89.73%
with recall values 85 and 92 for water and non water respectively at the patch
size of 16. Then our proposed network surpassed the performance of DenseNet
by reaching the test accuracy of 90.29% and recall values 86 and 93 for water
and non water respectively. Moreover, our experiments verified that neural
network were better than index-based approaches since the index-based approaches
did not perform well to extract riverbanks, small water bodies and
dried rivers. Qualitative analysis seconded the findings of quantitative analysis.
It was found that the proposed network was successful in creating attention
aware features of water pixels and diminishing urban, barren and non water
pixels.
All in all, it was concluded that the objectives of the research were met
successfully with the successful proposition of a new network.
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Index-Based Approach Deep Learning Convolutional Neural Networks Densely Convolutional Network Residual Attention Network State-of-the-art Approaches