Embedded Files
Exploring What’s Next 🚀🔍
Here’s a glimpse into the projects I’m currently working on. From advancing computer vision techniques to contributing to cutting-edge research in AI, remote sensing, and smart technologies, each project reflects my passion for pushing boundaries and solving real-world challenges. 🌍✨ I’m excited to share the journey as these projects evolve, inspire, and (hopefully!) make an impact.
Colorizing Syntetic Aperture Radar (SAR) Images
Colorizing Syntetic Aperture Radar (SAR) Images
While SAR images are great for flood mapping because they work in any weather or lighting, fusing them with multispectral images (MSI) improves results. The catch? Perfect co-registration is needed, which isn’t always practical. My research explores two questions:
While SAR images are great for flood mapping because they work in any weather or lighting, fusing them with multispectral images (MSI) improves results. The catch? Perfect co-registration is needed, which isn’t always practical. My research explores two questions:
🔎1) Do we really need RGB channels for colorization, or can other channels work?
🔎1) Do we really need RGB channels for colorization, or can other channels work?
🔎2) Can we skip the color guide altogether using Domain Adaptation?
🔎2) Can we skip the color guide altogether using Domain Adaptation?
🔎3) Is colorization even necessary for improving the accuracy of flood mapping?
🔎3) Is colorization even necessary for improving the accuracy of flood mapping?
Both the colorization and flood mapping are powered by AI techniques inspired by the UNet architecture.
This highlights UNet's versatility in handling multiple tasks!
The figure above is an illustration of the proposed framework.
The figure above is an illustration of the proposed framework.
Stage 1: composed of four convolutional layers (colored green) that utilize a ResNet34 backbone withinthe encoder, and four layers dedicated to upsampling (colored blue) within the decoder. The process of UDA is facilitated by the domain classifier through the multiplication of the domain loss gradient (L′d) with a factor of γ. Meanwhile, training for Ln continues in a standard manner.
Stage 1: composed of four convolutional layers (colored green) that utilize a ResNet34 backbone withinthe encoder, and four layers dedicated to upsampling (colored blue) within the decoder. The process of UDA is facilitated by the domain classifier through the multiplication of the domain loss gradient (L′d) with a factor of γ. Meanwhile, training for Ln continues in a standard manner.
Stage 2: consists of a simple 3D-UNet, which also consists of an encoder and decoder parts. Skip connections bridge encoder-decoder layers to enhance feature propagation.
Stage 2: consists of a simple 3D-UNet, which also consists of an encoder and decoder parts. Skip connections bridge encoder-decoder layers to enhance feature propagation.
In all the sample results, white = flood, black = no flood.
This result sample shows that the prediction of the colorized images is closer to the ground truth compared to the prediction from the non-colorized SAR images.
Notice how, when the majority of the image is covered by flood, the non-colorized images produce a lot of errors. In contrast, the colorized images successfully mitigate many of these errors, improving overall accuracy.
Cases where the flood doesn’t cover the majority of the image tend to produce more errors. The green boxes highlight areas where the colorized images reduce these errors, compared to the red boxes showing errors in the non-colorized images.
I was really excited when the idea for this project first came to me, and it’s easily one of the projects I’ve enjoyed the most. My colleagues and collaborators were incredibly supportive and eager to contribute to the idea. I truly enjoyed leading the project and getting involved in some of the programming work. The outcomes of the project were as follows:
I was really excited when the idea for this project first came to me, and it’s easily one of the projects I’ve enjoyed the most. My colleagues and collaborators were incredibly supportive and eager to contribute to the idea. I truly enjoyed leading the project and getting involved in some of the programming work. The outcomes of the project were as follows:
1) Colorization can be performed by channels other than RGB
2) Colorizing without the existence of a color guide (MSI) is possible with Domain Adaptation
3) Colorization does not necessarily improve the accuracy of flood mapping, as over-colorization can cause loss of information
This raises an important question: What is the optimal colorization threshold that preserves critical information without compromising the accuracy of flood mapping and other SAR applications? ➡️ this is the future direction of this research project. 🙂
Related Publications:
🔗N. Aburaed et al., "Multi-purpose Domain Adaptive Network for Unsupervised SAR Colorization," 32nd European Signal Processing Conference (EUSIPCO), Lyon, France, 2024, pp. 2222-2226, doi: 10.23919/EUSIPCO63174.2024.10715085.
🔗N. Aburaed, M. Al-Saad, M. S. Zitouni, M. Q. Alkhatib, S. Al Mansoori and H. Al Ahmad, "A Critical Examination of SAR Colorization Impact on Flood Mapping Accuracy," IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Athens, Greece, 2024, pp. 8504-8508, doi: 10.1109/IGARSS53475.2024.10642200.
Page updated
Google Sites
Report abuse