Hierarchical deep learning for better automatic pest classification and biodiversity monitoring in agroecosystems

Summary
This PhD project aims at developing novel algorithms and annotation tools for hierarchical classification, which has profound and immediate applications to both agriculture and ecology.

Context
Recently, remarkable innovations in deep learning have begun to transform agriculture – arguably, the most impactful human activity on earth. If used appropriately, machine learning and AI have the potential to drive a digital green transition that could make agriculture more sustainable, resilient and productive.

One increasingly important data-science task in agriculture is image classification, which is crucial for detecting, monitoring, and eventually mitigating the impact of weeds, pests and diseases. When applied to agricultural problems, image classification generally suffers from three important limitations: 1. low generalisability – i.e., it translates poorly to new contexts –; 2. data-hunger – i.e., large training datasets are required –; and 3. the lack of accessible implementations for practitioners.

As an alternative to the conventional “flat” classification, we propose to take advantage of the underlying hierarchical taxonomies of the predicted classes to constrain them in a priori trees, a currently underexplored area. As a simplistic illustration, instead of considering oranges, apples and lemons as independent, we may formulate the explicit tree: {{orange, lemon}, apple}, i.e. grouping citrus fruits, and embed this representation in the structure of a neural network.

Outcomes
We will use already available datasets to explore hierarchical classifications and test our hypotheses: increased prediction generalisability and network robustness as well as reducing the amount of data required. This project aims to produce both seminal machine-learning concepts and free and open-source tools for real-world problems.