Document Type

Thesis

Date of Degree Completion

Winter 2021

Degree Name

Master of Science (MS)

Department

Computational Science

Committee Chair

Răzvan Andonie

Second Committee Member

Boris Kovalerchuk

Third Committee Member

Szilard Vajda

Abstract

Nowadays, Synthetic Aperture Radar (SAR) images have been widely used in the industry and the scientific community for different remote sensing applications. The main advantage of SAR technology is that it can acquire images from nighttime since it does not require sunlight. Additionally, it can capture images under the cloud where the traditional optical sensor is limited. It is very convenient to use SAR image for surface water detection because the flatness of the calm water surface reflects off all the energy from the radar and this makes the surface water appears in a SAR image as dark pixels. The traditional way to mark out water from SAR images is by just using the thresholding method which a pixel is classified as water when its value is below a certain threshold. This method works fine in a plain rural area, but the complex features of urban areas make it more challenging, for example, highways and buildings’ shadows can be easily misclassified as water. To solve this problem, we propose the Fully Convolutional Neural Network (FCN) Encoder method, a deep learning model based on the convolutional implementation of sliding windows. The FCN Encoder is designed to detect water from SAR images by considering both the pixel intensity and the spatial information of the pixel (i.e., its neighborhood). In our experiments, we first train the network using the So2sat dataset, which contains patches of Sentinel-1 satellite SAR images. Next, we use the trained neural network to detect water from SAR images of several cities. The obtained results show satisfactory scores and also visually appear accurate. In the final optimization phase, we: a) train the FCN Encoder with our custom HARD dataset - a dataset with images that are harder to classify, and b) we optimize the hyperparameters of the model. We test the resulted classifier on public SAR images and compare it with other methods such as Smooth Labeling, Random Forest, and FCN Segmentation.

Share

COinS