In this thesis, a convolutional neural network (CNN) is investigated as a non-invasive approach for detecting drought stress in trays of Pinus Sylvestris seedlings using RGB imaging acquired in a controlled environment. The trays were divided into two classes, healthy and stressed, which were subjected to different growth conditions. Images of the trays were collected using an automated data acquisition protocol and were used to createa custom dataset for training a CNN through transfer learning.
Two iterations of data acquisition were performed, where the first dataset was used to develop and test the model’s performance, while the second was used to test the model’s ability to generalize to previously unseen data. To gain an insight into what the model bases its predictions on, saliency-based interpretability methods were applied to examine which regions of the image contributed most to its decision.Results from ten independent training runs on the first dataset show that the model reached an accuracy of 94% ± 1%, indicating that the model is able to distinguish between images of stressed and healthy trays. When evaluated on the second dataset, the model achieved an accuracy of 82%, suggesting that it is able to generalize to new data. These results indicate that the performance of the tested approach was successful in detecting drought stress under controlled conditions.
The result saliency maps suggested that the model focused more on background information rather than plant structure when making its prediction. While this does not take away the performance observed during testing, it highlights that evaluation metricsalone are not sufficient to evaluate a model’s performance, and emphasizes the importance of combining evaluation metrics with interpretability methods to understand thedecision-making nature of a CNN.