Lunchtime Chats with Kilian Murphy: Smart approaches to phase one habitat mapping using convolutional neural networks

Machine learning classification algorithms (e.g., Random Forests, Support Vector Machines, Convolutional Neural Networks) are models that can accurately predict on unseen data by learning from training data. In the case of phase one habitat surveying, these models can predict the fossitt code/habitat type at landscape scale from a representative sample of training data. When these methods are integrated with field data collected by expert surveyors it can produce high-resolution (~10m Setinel-2 multispectral satellite imagery) classification of all habitats in a project scope in a matter of hours, replacing the need for lengthy field site visits. Furthermore, resources saved on phase one habitat surveying can be re-allocated toward highly targeted and thus more concise and efficient site visits to ground-truth model predictions and re-train the model as needed to further increase the robustness of the model prediction. Working in tandem, this method unlocks new avenues for ecological consultancies to harness the power of deep learning technology to compliment traditional approaches and increase the robustness of their work.