DATA-310_Applied_Machine_Learning


Project maintained by cning0506 Hosted on GitHub Pages — Theme by mattgraham

Slice of Data Science - Caleb Robinson (Data Scientist at Microsoft AI For Good Research Lab)

February.11.2021

This Thurday, I attended the first episode of Slice of Data Science, which was a very fruitful event that I will continue to participate. It was very fortunate to meet Caleb Robinson, who is a Georgia Tech alum and a Data Scientist at AI for Good Research Lab. He started off by giving an overview of what AI for Good is, which focuses on the infusion of Data Science and AI to address the world’s greatest challenges. Specifically, Caleb and his team focuses on geospatial machine learning projects.

He is currently working on mapping concentrated animal feeding operations in the USA, partnered with Chesapeake Conservancy. The incentive behind the project is to address one of the Sustainable Development Goals (SDGs) - Goal 15: Life on Land. The goal aims to increase forest area as a proportion of total land area. Hence, Caleb’s key research objective is to identify what land is forest and what is not. He uses the technique of land mapping leading to a high-resolution land cover map. He then explains the reason behind the use of machine learning. First, manual labeling is expensive. Second, the low efficiency. With high cost and low efficiency of labeling, it is crucial to train models to assist the progress.

The application goal is to create an accurate, high-resoltuion land cover map. The process begins with a high resolution input, which is used for the training of convolutional neural network, resulted in high-resolution predictions. The research goal is to integrate low-resolution data sources. The product of this research is a large scale high-resolution land cover mapping with multi-resolution data. Therefore, the team developed super-resolution and uses other data fusion methods that improve geographic generalization.

As a result, he presented a demo of the interactive segmentation tool that is named under Microsoft AI for Earth. The user can retrain the model by transfer learning. He also talked about he most recent project on mapping polutry barns and how he identify concentrated animal feeding operations (CAFOs) at scale. Last but not least, he introduced the Machine Learning competition - 2021 IEEE GRSS Data Fusion Contest Geospatial AI for Social Good. The challenge of the competition is to identify land cover change at 1 a 1m resolution. We only have access to 30m resolution labels and the labels are noisy!