Understanding Compositional Targets in the Machine Learning Context

The dairy industry uses perennial ryegrass as fodder for cows. Optimising herbage production and effective grass utilisation is essential for achieving profitability in dairy farming. Accurate estimation of pasture biomass is essential for efficient farm management. Traditionally, this estimation involves labour-intensive and often destructive methods, limiting its scalability and precision. Computer vision and deep learning provide an effective solution for automating the non-destructive estimation of pasture biomass from proximal imagery of paddocks at scale and speed. This thesis presents a novel deep learning approach for biomass proportion estimation directly from proximal paddock images, demonstrating superior performance compared to current state-of-the-art methods. It shows promise in replacing labour-intensive manual estimation or other destructive techniques. The thesis also investigates the transferability of the deep learning model across different geographies, concluding that localised training provides better results than geographical transfer of models across regions with limited data. Predicting dry biomass proportions is a case of multi-target regression within machine learning. An image of a small quadrat from a paddock often includes different components of plant species (grass, clover and weeds) in varying proportions. The ground truth for the dry biomass of each component is represented as proportions summing to a total of 1. This data presents a compositional structure where the targets are interdependent due to the constant sum constraint. We introduce the term Compositional Targets (CoTa) to denote this type of data that are the targets for prediction. CoTa poses unique challenges for neural networks in a multi-target prediction context. The constant sum constraint of CoTa causes the scale of prediction errors for the components to be disproportionate. Therefore, predicting CoTa requires a distinct approach compared to traditional multi-target regression problems, where targets may be interdependent but not constrained. Although Compositional Data Analysis (CODA) approaches from statistics have been used extensively for analysing compositions, we show that directly adapting the CODA technique of logratio transformation of targets to handle compositional data for deep learning does not outperform the standard deep learning approach for multi-target regression with CoTa. This thesis introduces the CoTaNET framework, a specialised neural network framework for predicting compositional targets. The CoTaNET framework is developed through comprehensive experiments on various datasets that exhibit compositional target structures. It features appropriate activation and loss functions suitable for this class of data, offering a significant advancement in the field. The main contributions arising from the work described in this thesis are the development of a deep learning approach for biomass proportion estimation that outperforms existing methods, the identification of the limitations of transferring models across different geographies, the exploration of the application of statistical techniques from Compositional Data Analysis (CODA) in a deep learning context, and the creation of the CoTaNET framework, which provides an effective solution for predicting compositional targets across various datasets.