Multi-modal fusion for remote biodiversity surveys of linear green infrastructure

Scientific background

The agricultural sector has a huge role to play in conserving biodiversity and mitigating its impact on the climate crisis. Biodiversity is an important indicator of the overall health of a habitat, and the environment in general. However, agricultural activities, typically focused on growing a single crop type over a large surface area and reliant on a number of substances (ranging from fertilisers to pesticides and herbicides) that are known to cause biodiversity losses, not just in arable fields, but also in the areas surrounding them.

This project proposes to develop the means of automatically monitoring the biodiversity of habitats surrounding cultivated land, such as hedgerows, using computer vision techniques that incorporate information from several input sources including satellite, UAVs, and mobile robots. Research methodology

The project will initially focus on the continuation of previous work applied to road verges, where a computer vision approach (based on deep convolutional neural networks) will be extended as follows: (1) re-training of the model using data supplied by industry collaborator Gaist,

(2) transfer of knowledge learned on road verges to other domains, such as hedgerows and additional geographic locations, (3) incorporation of hierarchical and ordinal relationships when classifying images.

The second phase of the project will focus on incorporating additional imaging modalities, such as satellite imagery, UAV photographs, and images collected by mobile robots. In the third and final phase, a hierarchical classification system that effectively utilises information from each of the different modalities.

Training

The PhD candidate will have the opportunity to study and develop computer vision techniques and contribute to the state-of-the-art in deep learning-based methods (i.e., convolutional neural networks, vision transformers). As such they will develop skills in image processing, machine learning, and data visualisation. The candidate will also develop skills in working with several hardware platforms, and operation of a UAV.

They will develop their oral and written communications skills through production of journal manuscripts, conference and seminar talks, poster presentations, and engagement with collaborators from other disciplines.