Noah Dewar

Features used to train the convolution neural nets and the net architecture used.

As part of a class project, with my teammate Aakash Ahamed, we explored the use of various machine learning models in order to predict year-end county level agricultural crop production from satellite imagery taken during the summer months. A wide variety of machine learning models (Naive Bayes, random forest, feedforward neural nets, etc.) were tested but convolution neural nets provided the best performance. The median test accuracy achieved with convolution neural nets was ~91%.

The image on the left and right show labelled training data ready for a convolution neural net. The examples on the left are of center pivot irrigation, while those on the right are examples without center pivot irrigation.

Managing groundwater sustainability in the face of climate change is a global challenge that has serious implications for food security in both developed and developing nations. One challenge facing sustainable management in developing nations is a lack of data on how much and where groundwater is being used. Transporting groundwater requires infrastructure, which can be expensive. Thus, irrigated agriculture is often grown in a circular pattern around wells, this is known as center pivot irrigation. One potential way to monitor groundwater usage around the world is to use satellite based remote sensing and machine learning to identify active center pivot irrigation and use vegetation health as a proxy for groundwater extracted.

As part of a hackathon held at Stanford University that was focused on big data and water, I investigated multiple methods of automatically detecting center pivot irrigation in satellite images. While my team was developing a machine learning approach using convolution neural nets, due to the time constraints of the hackathon, simpler methods such as Hough transforms ended up providing better results.

I graduated from McGill University in Montreal with a joint major in physics and geophysics in 2012. After graduating I worked in ground-based geophysical surveying with Quantec Geosciences and then in airborne geophysical surveying with Sander Geophysics Limited. After travelling the world and running a variety of natural resource related exploration projects, I felt stagnant. I had learned all I could in my current position. I started applying to graduate programs while living in Barranquilla, Colombia and running an airborne magnetic and gravimetric survey. I was accepted to Stanford University in 2016 and immediately set about taking classes and learning everything I could while at the same time applying what I had learned in industry to the problems my research group was trying to solve.

I graduated from Stanford June 2021, during the height of the pandemic. During the end of my PhD I launched a specialized geophysics and machine learning consulting shop, Dewar.AI. After graduating I worked as a data scientist for a little while but soon realized I needed a new mission. In May 2022 I found that mission when I joined my good friend Adrian Santiago Tate as cofounder and chief technology officer of HighTide Intelligence, a flood risk startup with the impossible seeming mission of solving flooding.

Highlights from my CV: Graduated from McGill University in 2012 with a 3.51 GPA and a joint major in physics and geophysics. Three years of experience as a geophysical analyst and field crew chief with Sanders Geophysics Limited. Graduated from Stanford University with a doctorate in environmental geophysics with a 3.87 GPA. Fluent in English, Matlab, and Python. Competent in Spanish and C++. Years of leadership and coaching experience in club ultimate frisbee teams at both McGill and Stanford. Vice president of the Stanford chapter of the society of exploration geophysicists for three years in a row. Student lead on the near surface geophysics technical subcommittee for the society of exploration geophysics. Stanford classes include machine learning, computer vision, numeric methods, and deep learning with convolution neural networks. McGill classes include calculus, advanced calculus, ordinary and partial differential equations, and linear algebra.