Scivision

Applications

Repurposing MapReader model for Automated Plant Phenotyping

Understanding how the genetics of plants interact with their environment to produce certain characteristics or 'phenotypes' is critical to understand how they might grow under different conditions. In relation to agriculture, extracting accurate data on phenotype may help us to better manage plants to produce higher-yield, more resilient crops, and plan for future food security by predicting how crops may grow under various climate change scenarios.

State-of-the-art plant phenotyping platforms have recently been established in the UK, such as the National Plant Phenomics Centre that collect high spatiotemporal resolution imagery of plants, as well as data on plant genetics and environmental conditions. However, extracting phenomic data from these images is expensive and time consuming to carry out manually.

We are working with the National Plant Phenomics Centre (NPPC) to automate extraction of plant phenotype data from various datasets, including time-series images of individual Brassica napus plants. We want to track the change and emergence of different plant structures (such as leaves, flowers, branches and seed pods) over time.

Through the Scivision framework we were able to load example plant data (at individual and satellite scale) and perform inference with a MapReader model trained on NPPC images of individual Brassica plants.

MapReader is an end-to-end computer vision (CV) pipeline developed to analyse historical maps as scale, using a patch classification approach to identify landscape features i.e. railways and buildings, and examine change in these classes over time. Through the Scivision framework we were able to use an adapted version of a model trained through the MapReader pipeline to classify various plant structures (i.e. flowers, leaves, seed pods) at patch level, and examined how the abundance and distribution of these different structures change over time as plants grow.