Deep Learning for Automated Discovery and Analysis of Plant Leaf

2016_25: Deep Learning for Automated Discovery
and Analysis of Plant Leaf Physiographic
Characters Related to Climate Change
Supervisors: Dr Ben Glocker ([email protected]), Professor Norman
MacLeod (Earth Sciences/The Natural History Museum)
Department: Computing
Characteristics of angiosperm leaves (e.g., size, aspects of shape) vary to reflect local
climate states. While procedures exist to quantify this relation and predict past states
from leaf data, these methods suffer from lack of automation, limits on the number of
used features, and reliance on linear data-analysis procedures. Hence, those
procedures may not be sufficient for modelling the entirety of a complex, non-linear
relationship between leaf characteristics and climate state. In this project, we explore
cutting-edge machine learning methods, and in particular, deep learning techniques,
that have been shown to be capable of capturing complex relationships in highdimensional data, and are now considered state-of-the-art in many applications such
as object recognition and natural language processing. Deep learning methods are
attractive as they can take raw data (such as images) as input and feature learning is
part of an end-to-end optimisation procedure. While more low-level features are
extracted on the first few layers of a deep neural network, high-level, non-obvious
features often emerge in deeper layers. This project has the potential to reveal
unknown relations between the input, i.e. raw photographs of leafs, and the output,
i.e. climate state that could be used both to improve estimates of climate change from
an automated analysis of leaf form, and to estimate plant response(s) to climate
change. We hypothesise that a deep learning approach can yield to more accurate
predictions compared to current procedures, such as CLAMP, in which only a few
handcrafted features based on expert knowledge are used for prediction.
We intend to bring a multi-disciplinary team together to supervise the student in their
construction of a new approach to inferring ancient and modern climate states from
leaf physiognomic data. Currently most widely used leaf-based climate inference
procedure employs manual scoring of the leaf character states by experienced human
technicians to collect its data. The student will extend and automate this process using
deep learning methods to comprehensively explore the relation between leaf form and
climate variation to (1) better understand the manner in which plant leaves respond to
climate change and (2) develop leaf geometric proxies for the inference of climate
state from leaf physiographic data. The accuracy of this new approach to climate
affect/inference will then be tested against estimates derived from current semiquantitative approaches to this problem.
For more information on how to apply visit us at
Science and Solutions for a Changing Planet
A large quantity of plant leaf specimens and plant leaf images currently resides in the
collections of The Natural History Museum and in the collections of the co-supervisor’s
research colleagues. Modern climate data, of course, is available from a wide variety
of sources.
The supervision team includes an expert in image analysis and computer vision with
a focus on semantic understanding of images using machine learning (Ben Glocker)
and an expert in morphometrics, leaf physiography, palaeontology, and natural history
applications of statistics with a specific track record of publications on the
morphometric analysis of leaf form (Norman MacLeod).
For more information on how to apply visit us at