Last August I started a PhD residency at . X is a research and development branch of Google (or more precisely Alphabet). is a diverse group of inventors and entrepreneurs who build and launch technologies that aim to improve the lives of millions, even billions, of people. The goal is 10x impact on the world’s most intractable problem. X approaches projects that have the aspiration and riskiness of research with the speed and ambition of a startup. 

In 2010, Google founders Larry Page and Sergey Brin decided to form a new division of the company to work on moonshots: far-out, sci-fi sounding technologies that could one day make the world a radically better place. It was a grand experiment with ambiguous wording in vision. 10 years in, X website mentions that it has incubated hundreds of different moonshot projects, many of which have gone on to become independent businesses. 

As an AI Resident, I'm engaged in an early stage project where I work on various facets of GeoAI model development and deployment, encompassing everything from model design to DevOps and MLOps practices. This role has afforded me the great opportunity of learning the implementation and deployment of AI models at the forefront of the industry, while also being exposed to the practical, conceptual, and theoretical challenges that emerge due to scalability issues.

My academic background and research in the Geography Department at ÃÛÌÒ´«Ã½Æƽâ°æÏÂÔØ have been instrumental in preparing me for this role. In particular, courses such as Geospatial machine learning in the Geography Department, along with deep learning courses in the Computer Science and Applied Mathematics Departments, have been significant. Additionally, an equally impactful course, software engineering for scientists, has also greatly contributed to the skills I use on a daily basis in this position.

In light of this experience, I look forward to my return to academia to complete my PhD leveraging the insights and skills I have gained in this role. I am excited to integrate the practical knowledge I have acquired from the industry into my doctoral research.

Behzad Vahedi page

Off

This past February, I received the Outstanding Student Presentation Award (OSPA) from the American Geophysical Union (AGU) for his presentation at the AGU Fall Meeting 2021. According to the AGU, "this honor is awarded for only the most exceptional presentations during AGU Fall Meeting 2021."

I am a PhD student at our department and prior to joining CU, I completed a B.Sc. in Geomatics Engineering and an M.Sc. in GIS Engineering at K.N.Toosi University of Tech. in Iran, and an M.A. in Geography at University of California Santa Barbara. My research focuses on spatiotemporal machine learning, including both classification and regression problems and my work is motivated by two major application domains: the classification of sea ice from satellite-based radar imagery captured over the Arctic, and forecasting the spread of COVID-19 across the United States.

I won the OSPA award for a presentation titled "A Comparison of Classic Deep Learning Architectures For Sea Ice Classification From SAR". In this work, I explored the idea of transfer learning in the context of classification of different sea ice types from Sentinel-1 Synthetic Aperture Radar (SAR) imagery. The state-of-the-art image classification models are trained on images of everyday objects such as animals, symbols, and vehicles. Needless to say, this is very different from the contents of a standard optical or radar image acquired by a satellite. The idea of modifying a model trained on everyday objects in order to use it in a different domain, remotely sensed radar imagery in this case, is called transfer learning.

I compared the performance of some of the most used models from the field of computer vision in the task of classifying sea ice in radar images. To this end, I examined whether what these models have learned from everyday images can be transferred to the radar imagery, and if so, to what extent. This is significant because the amount of labeled radar imagery is limited. The findings of this study can help researchers make informed decisions when designing deep learning-based classification models for sea ice. 

Inspired by these findings, I am currently working on developing more powerful classification models that can achieve high performance with limited data and hopefully help those experts in the task of sea ice charting in the near future.

I would like to thank his advisor Dr. Morteza Karimzadeh and his co-author Dr. Benjamin Lucas for their help and support, and the ÃÛÌÒ´«Ã½Æƽâ°æÏÂÔØ Geography for the great environment it has provided for research and collaboration. 

Off

Presentation Abstract

During the last decade, advances in the state-of-the-art deep learning models, in particular convolutional neural networks, have facilitated significant improvements in image recognition tasks. In fact, on the benchmark ImageNet dataset, the state of the art is now recognized as performing better than human. As a result, many adjacent tasks, including image recognition in remote sensing, have adopted these state-of-the-art models with little investigation into their transferability. For instance, the common image datasets—from which pre-trained model weights are derived or modern architectures are evaluated on—contain R-G-B images of everyday items such as animals, symbols, and vehicles. Needless to say, this is very different from the contents of a standard optical or radar image acquired by a satellite.

In this work we explore this idea of transferability in the context of the classification of sea ice type from Sentinel-1 SAR imagery. We compare the performance of a basic CNN with 4 significant models from the field of computer vision—AlexNet, ResNet, VGG, and Inception—in the task of classifying sea ice in a region of the Chukchi Sea, a sea of the Arctic Ocean. We extend these experiments further to compare models that have been pre-trained on the ImageNet dataset with models where the parameters are randomly initialized, to demonstrate whether pre-trained models are beneficial for this application. The performance of models is compared using overall accuracy and F-1 score. Finally, we hypothesize why some models perform better on our dataset than the others, and we conclude by explaining how the results inform our model choice for future sea ice classification projects.

Off

Off