1. About Us
The Brisbane Flora and Fauna Society (BFFS) is a not-for-profit committed to the preservation,
protection, and enhancement of Brisbane’s diverse ecosystem. BFFS is composed of dedicated
environmental enthusiasts, biologists, and volunteers who work tirelessly to safeguard the
region’s rich biodiversity, focusing specifically on the city’s unique flora and fauna. Through
education, conservation initiatives, community participation, and collaboration with government
and non-government entities, BFFS is a frontline guardian of Brisbane’s natural heritage,
balancing urban development with ecological sustainability.
2. Project Motivation
We currently use trained experts to monitor the growth of different plant species in Brisbane
parklands. These experts label and count the ‘healthy’ plants in the parks – these are plants
that encourage biodiversity. The experts also identify and remove ‘weeds’ – plants that grow too
quickly and kill other healthy plants. This existing approach works well, but is difficult to scale as
training experts is expensive and time-consuming.
We want to trial a new approach where non-expert volunteers upload a photo from a park and a
ML algorithm classifies the species of the plant. If the photo is classified as a healthy plant, we
will use that to monitor the population of that specific species. If the photo is classified as a
weed, the volunteer will mark that plant for future removal or weed-spraying.
3. Project Description
We want you to complete a study into the feasibility of using AI to classify plant species. We
have collected data for 5 different healthy species and 5 different weed species.
Figure 1. Healthy plant species.
1
Note: Brisbane Flora and Fauna Society is fictitious – logo and ‘About Us’ have been generated with AI.
Figure 2. Weed species.
We provide you with a dataset folder, where you will find:
● labelled – this folder contains example images from the 10 different species, sorted into
individual species folders. We have used one of our trained experts to label this data.
There are between 40-70 labelled images for each plant species. There are a total of
548 images.
● unlabelled – this folder contains approximately 430 extra images of plant species that
have not been labelled by an expert. The majority of these images should belong to one
of the provided 10 species, and there may be a small number that do not belong to any
of the 10 species.
We additionally have a held-out set of 300 images, with 30 images per plant species, that we
will use to evaluate the AI method you design. We will not release these images to you, but we
will test your final solution on this data to measure the reliability of your method.
4. Project Deliverables
We have a number of constraints we require you to follow in your investigation.
Constraint 1: The model
For classifying images, use the Pytorch implementation of a ResNet18 architecture
2
initialised
with the available pre-trained weights from training on ImageNet. You may change the final layer
of this model to be suitable for this dataset with 10 classes.
Constraint 2: Data subsets
You should create a training and validation subset from the data you have been provided. Each
plant species must have at least 20 images in the validation subset. When presenting your
results, you should report performance on your validation subset.
Using these constraints, we would like you to investigate and present results for the following
tasks. We will meet with you in Week 7 for 10 minutes. You should bring printed material to
communicate your findings with us, and use the meeting to explain the findings and answer our
questions.
2
See https://pytorch.org/vision/main/models.html
Task 1: What performance can you achieve with only the labelled data?
You should be prepared to discuss and share materials showing the following:
● Statistics about the training and validation dataset (e.g. number of images, total and per
class).
● Hyperparameters used during training the model.
● Training and validation loss and performance curves.
● Final performance including metric/s and a confusion matrix.
● Any other useful visualisations or qualitative results.
Task 2: How can you utilise the unlabelled data as well as the labelled
data? How does this affect training and performance?
You should be prepared to discuss and share materials showing the following:
● How you used the unlabelled dataset.
● Any relevant visualisations or plots you create based on this experiment.
● Final performance including metric/s and a confusion matrix.
Task 3: Given your investigation, do you have any recommendations for
us?
This may include recommendations relating to:
● Are the models performing at a suitable level for our described use-case?
● How could we create a model with better performance?
● Do you have any concerns about the data we have collected?
● Any other recommendations you may have.
Where suitable, a recommendation should be supported by the material brought with you to the
meeting.
Task 4: Provide us with your recommended ML algorithm.
Using the provided submission link, share your recommended ML algorithm with us, including
the architecture and trained parameters. We will test this algorithm on our held-out test dataset
to understand the reliability of your algorithm.
6. Project Deadline
We look forward to meeting with you in Week 7. We will reach out to you in the coming weeks to
schedule a 10 minute meeting, where you can present your findings to one of our
representatives.
程序代写代做C/C++/JAVA/安卓/PYTHON/留学生/PHP/APP开发/MATLAB
本网站支持淘宝 支付宝 微信支付 paypal等等交易。如果不放心可以用淘宝交易！
E-mail: itcsdx@outlook.com 微信:itcsdx
如果您使用手机请先保存二维码，微信识别。如果用电脑，直接掏出手机果断扫描。
