University of Twente Student Theses
Integrating Convolutional Neural Networks and Conditional Random Field Inferencefor Land-Cover Classification of Multi-Scale Remote-Sensing Images
Gupta, Aman (2018) Integrating Convolutional Neural Networks and Conditional Random Field Inferencefor Land-Cover Classification of Multi-Scale Remote-Sensing Images.
|
PDF
4MB |
| Abstract: | The abundant availability of multi-scale very high-resolution (VHR) remote-sensing imagery precipitated the need for an image classification method to automatically learn relevant features from raw images and make land-cover class predictions in an end-to-end framework. For this, deep learning methodologies were adapted, specifically in the form of convolutional neural networks or CNNs. In this research, we investigate the ability of CNN structures to dynamically fuse multi-scale VHR satellite images, taken by the WorldViewIII sensor over parts of Quezon City, Mexico, to produce high-accuracy land-cover classification maps. We do so by comparing the classification results of a CNN that is designed to fuse discrete sets of multi-scale images, with the results of a CNN that was trained on images that were pan-sharpened using an independent state-of the-art technique. In addition, this research applies the mean-field approximation to the dense-CRF inference which models label and spectral compatibility in the pixel-neighbourhood effecting a more locally, smooth segmentation. Finally, a recurrent neural network or RNN is also implemented in order to enhance the stand-alone fusion network by accounting for local class-label dependencies within its end-to-end structural framework by concatenating the class-probability scores of the first FCN with the raw input of the following FCN instance. We compare the dense-CRF optimized classification results with those obtainedby applying the trained RNN classifier. Classification results indicate that our designed classification FCN models have performed favourably. The fusion network consistently produces high-accuracy land-cover classification maps effectively demonstrating the ability of FCN structures to dynamically fuse multi-scale images. Furthermore, the RNN results show an improvement on the singular fusion network performance. This illustrates the capacity of FCN architectures to account for local spatial and spectral dependencies in an image space, when implemented as a RNN |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/85866 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page