Clump It Up!

Clump It Up! was a course project for CMPT 898: Automated Agricultural Analysis, part of my M.Sc course requirements. The course pairs together students from Computer Science with students from Agriculture to try to solve a problem related to automating analysis of agricultural data. These projects contribute to the P2IRC Project at the University of Saskatchewan. For our project, we worked on automatically counting emergence of brassica carinata plants in a field. We obtained a dataset of images taken from the Protractor System which drove across a field, taking images as it went.

We decided that it would be easier to count plants in a small area of the whole image, rather than trying to train a network to count all the plants in a large image. To accomplish this, we devised a pipeline that was made up of two steps: detection and cropping of plant “clumps”, and counting of plants within a single “clump”. We tried a few different methods for each step. To accomplish detection of “clumps” we applied an Excess Green Threshold, Faster RCNN, the YOLO architecture, and a U-Net architecture. For counting of plants within a clump, we trained some simple machine learning classifiers, namely K-Nearest Neighbours, Support Vector Machine, and Gaussian Naive Bayes. We also attempted counting using a basic CNN architecure and a Capsule Network. The multitude of approaches gave each member of the group plenty to work on, and gave us a unique opportunity to compare some different approaches to a single problem.

My contributions to the project were managing the dataset and coordinating annotations, documenting the project, and training and testing some simple encoder-decoder networks that directly output the number of plants in a given image. I also presented a poster about our project at the 4th annual P2IRC symposium, in Saskatoon SK, which won the best poster award.

Check out the Repository