Using satellite imagery to monitor seaweed biomass
The fucoid Ascophyllum nodosum (Ascophyllum) is commercially the most important species to the Irish seaweed industry. Today, the harvesting of Ascophyllum occurs on an industrial scale throughout Galway, Mayo and Donegal, and to a much smaller extent in other western counties such as Clare, Kerry and Sligo.
The challenge for the burgeoning seaweed industry in the west of Ireland is to develop technologies to accurately define the biomass of Ascophyllum to ensure the sustainability of the resource. The effective monitoring of Ascophyllum is crucial to ensuring the sustainability of the Irish seaweed industry, particularly considering the ever-increasing popularity of seaweed and seaweed-based products.
The use of satellite Earth Observation (EO) techniques have arisen as a powerful tool for the monitoring and classification of seaweeds. One of the clear benefits of EO is its minimal, or non-existent, operating cost for users and periodical acquisition of images without the need of any operator.
The objective was to create a platform that would allow users to easily and quickly get seaweed biomass information from their satellite images. As with any of our projects, the aim was to create a platform that was easy to use and with little to no friction for the end user.
A web application was created that allows users to supply satellite images, that are then processed by our pipeline. The pipeline transforms the image and extracts features from it, which are then used to make a prediction of the seaweed biomass. The pipeline then produces a GeoTIFF image in the form of a heat map representation of the seaweed biomass and statistics related to the biomass in the image.
The biomass predictions are done with the use of machine learning. Our machine learning model is a Convolutional Neural Network (CNN), which is commonly used to identify patterns in images, which is perfect for this type of project.
The backend, like many of our other projects was built using Amazon Web Services (AWS). The platform was created with an API first approach. We created serverless APIs with the help of AWS Lambda and API Gateway.
We made use of AWS Sagemaker to help us train, build and deploy our machine learning model. We created a pipeline around the deployed model, to allow us to process data before and after a prediction. AWS Step Functions was used to connect multiple lambda functions, forming the pipeline. The pipeline would extract data from the satellite image and pass these details to our machine learning model for prediction. These predictions are then transformed into the heat map GeoTiff file. The lambda functions that make up this pipeline were written in Python, to allow us to make use of popular data analysis libraries.
As satellite images can be very detailed, and therefore, very large, we made use of Elastic File System (EFS) volumes attached to our lambda functions. This gave us far more room than the 512mb /tmp directory provided with each lambda function. In order to make use of EFS volumes we also set up a VPC.
We created three web apps using React, all of which were built as progressive web apps (PWAs). We made use of storybook.js to help us build and test our React components.