Cloud Hosting

Ever since the industrial chemist Leo Baekeland began synthesizing phenol and formaldehyde in 1907, the world has developed a love-hate relationship with the resulting polymer: plastic.

While plastic is convenient, durable, and cheap, 50% of all plastics (about 150 million tons every year, worldwide) are used only once and then thrown away. Even for those who dutifully recycle our plastic water bottles and sandwich bags, were only tackling a small part of the problem. Thats because heavy winds and rain carry huge amounts of plastic waste along city streets and into the stormwater system, where it likely flows directly into creeks, rivers, bays, and eventually the ocean, with no treatment to filter out plastics.

Ora Loma Marsh in Hayward, California.

Considering the size of the problem, theres relatively limited infrastructure in place to capture and treat stormwater, says Tony Hale, program director for environmental informatics at the nonprofit San Francisco Estuary Institute (SFEI).

Thats where SFEI is looking to use research and dataand most recently, dronesto make a difference.

In addition to sending out crews of people on foot to count and collect trash in local waterways, SFEI began using camera-equipped drones to assess that waste on a much larger scale.

Most ground crews working for stormwater programs monitor trash once a year, twice if were lucky, Hale says. So what we can learn about trash and its impact on communities is limited by the number of people we can afford to send out.

With drone photography, we can track all of the trash in a creek, river, or stream, examine how its distributed, and then apply machine-learning algorithms to analyze those images as often as we want, Hale says.

The drone research is part of a new project by SFEI and its sister organization Southern California Coastal Water Research Project, through funding from the Ocean Protection Council, to validate trash-monitoring methods, and produce a trash-monitoring playbook that community cleanup groups, municipal programs, environmental agencies, and ecologists can learn from and put to use. The effort studies initiatives such as plastic bag bans to urban rain gardens.

Our mission is to help city planners find the best ways to filter their stormwater and stop contaminants such as trash and plastics from entering their protected wetlands and public waterways, Hale says.

Deep-Learning Cleanup Crew

By sending drones over the San Francisco Bay and neighboring tributaries, SFEI collected some 35,000 images in its initial foray.

Covering so much ground so quickly was amazing, Hale reflects. But his excitement soon faded, as the reality of crunching so much data in a reasonable amount of time set in: It took us almost a month to process these images.

Using 2,000 annotations to describe various trash particles, Hale and his team were training an open-source TensorFlow machine-learning algorithm to identify the type, quantity, and location of each particle of trash depicted in those 35,000 images.

To speed up the analysis, SFEI partnered with Kinetica, a data analytics startup that participates in the Oracle for Startups program. It put SFEIs trash-detection model into a Docker container and then brought it into Kineticas active analytics workbench, says Kinetica CMO Daniel Raskin. Using a Python API, Kinetica then streamed the images into a table, where they could be stored, categorized, and labeled.

Were not just ingesting these images and distributing them inside our platform Raskin says. Were also running SFEIs trash-detection model to classify all of the images as they hit our database.

This gives SFEI more than just a giant image catalog. The California water-quality watchdog can now visualize each of the 35,000 images based on its geographical location and trash profile.

Initially, Kinetica ran SFEIs deployment from a distributed CPU framework, on its own 4-core machine, using managed Kubernetes. It took us about 10 days to run the entire simulation, says Nick Alonso, a solution engineer at Kinetica who works on the SFEI project.Even after moving the application to a server using a single GPUprocessors that are well suited to machine learning workthe simulation still took the better part of a week.

Kinetica then decided to run SFEIs entire workload on Oracle Cloud Infrastructure, using eight V100 GPUs. Were no longer talking about days to run this simulation, Alonso says. Were doing it in hoursabout 18 hours and 26 minutes, to be exact.

See the original post here:

Drones And Artificial Intelligence Help Combat The San Francisco Bays Trash Problem - Forbes