Out-Of-Distribution Weed Detection and generalization

Weeds can negatively impact agricultural fields, which compete with crops for nutrients, water, light, and space. Herbicides are still the most used control method. However, over time, this strategy may negatively impact the environment, and organic farming standards forbid the excessive use of herbicides. For efficiency, drones collect images from fields to automatically sense and detect weed locations to spray the infected area. However, the weather, air pressure, and time of the collection would impact the drone flight, resulting in blurry images with different background lighting, making it more challenging for deep neural networks to detect weeds with little training data. Moreover, weeds can grow with different growth rate stages and various types, making it harder to disentangle them from the crops. For that reason, in our work, we will train our network to minimize the empirical risk across different environments using data augmentations and various trainset distributions. Finally, our main objective is to create a generalizable framework that can generalize on different environments besides detecting hard OODs to update the model later.

Faculty Supervisor:

Liam Paull

Student:

Partner:

Precision AI

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

Université de Montréal

Program: