TERRAIN CLASSIFICATION BY USING HYPERSPECTRAL AND LIDAR DATA

Date

2023-6-14

Author

Peker, Ali Gökalp

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

272
views

274
downloads

Cite This

Effective modeling of multiple sensor data is an important research direction towards improving the state-of-the-art in remote sensing. In this thesis, we address this issue starting from machine learning-based classification pipelines and propose a deep learning-based terrain classification framework for hyperspectral and LiDAR data that does not require sensor-based model specialization with a new building block for detecting spatial and spectral patterns. This framework enables models that can outperform existing classification methods. Deep learning models still encounter significant difficulties in regions covered with shadows. To address this issue, this study proposes a data augmentation approach based on generative adversarial networks (GANs) and a novel loss function that combines the transitive style transformations and unpaired matchings with correlated samples. This novel loss function leads generation of synthetic samples for regions masked by building or cloud shadows, thereby boosting the performance of deep networks in recognizing such regions. Qualitative and quantitative evaluations show that the proposed methodologies can be used to build models that can fuse multi-sensor data and improve classification results under shadowed regions.

Subject Keywords

Data fusion, generative adversarial networks, deep neural network, hyper spectral imaging, light detection and ranging, Data fusion, generative adversarial networks, deep neural network, hyperspectral imaging, light detection and ranging

URI

https://hdl.handle.net/11511/104448

Collections

Graduate School of Informatics, Thesis