RAYN Vision System Analytics

RAYN Vision System Analytics is the analysis module of the RAYN Vision System, and works as a tool in its own right: it reads hyper- and multispectral images from RVS Cameras on the same network, but equally from any other imaging source. It turns them into numbers: shape and size parameters, and a set of reflectance indices computed per region of interest.

The work happens in a graphical interface rather than a notebook. Regions of interest are drawn directly on the image, masking methods and thresholds are chosen interactively so the plant can be separated from its background, and the result can be previewed before a run is committed. Those are the standard options; underneath, the analysis is a Python script. A base script ships with the application, and users can add their own analysis and mask scripts. A JSON config file that accompanies a script declares which of its settings and functions appear in the interface, so a custom pipeline arrives with its own controls instead of hardcoded parameters — no changes to the application itself.

My Role

I co-developed the application with a colleague, who took on the user interface and the application core while I worked on the analysis and customization side. That meant the analysis and mask scripts — the Python deciding how a plant is separated from its background and which indices are computed from what remains — and the config-driven mechanism that lets a script bring its own settings and functions into the interface. Where the functionality we needed was missing in PlantCV, I contributed it upstream rather than reimplementing it privately: improving a wheel instead of building another one.

Alongside that I wrote specifications for features and validated the application against real experiments, checking that what came out was correct and usable for answering a research question. I also worked on the data pipelines and the visualizations that sit at the end of them, since a table of reflectance indices is not yet an answer — researchers need to see the result before they trust it. Working on the camera in parallel meant I could keep the two halves aligned, from what the camera writes to what the analysis expects to read.

Open by Design

The application is built on PlantCV, the open-source image analysis package for plant phenotyping, and is itself open source under the Apache 2.0 license, with the source and installers on GitHub for Windows and macOS. That was a deliberate position rather than an afterthought: a community-driven analysis pipeline is not something to put behind a license key, and researchers need to be able to read the code that produced the numbers in their paper.