Masterarbeit
Detection of Water Stress in Plants by Image Recognition
Completion
2025/09
Research Area
Students
Anubhav Saha
Advisers
Prof. Dr.-Ing. habil. Stefan Streif
Dr.-Ing. Sebastian Heil
Description
Early detection of water stress in plants is essential for improving crop management and ensuring sustainable agricultural practices. The objective of this study is an investigation of deep learning techniques for both image classification and object detection to identify water-stressed plants. A dedicated dataset of lettuce plants was collected under controlled environmental conditions, comprising both healthy and water-stressed specimens at varying growth stages. Images were captured under consistent lighting and camera settings to reduce noise, while data augmentation techniques—such as rotation, flipping, zooming, and cropping—were employed to enhance model generalisation and mitigate the limited dataset size.
Several convolutional neural network (CNN) architectures, including VGG16, ResNet50, EfficientNet, MobileNet, and DenseNet, were evaluated for classification tasks. Among these, EfficientNet-B0 and DenseNet-169 achieved the highest accuracy of about 95%, which is attributed to their balanced architectural depth and efficient feature reuse. Larger and more complex networks exhibited overfitting tendencies, likely due to the relatively small dataset and limited intra-class variability. To address more realistic scenarios involving multiple plants per image, YOLO-based object detection models were implemented. Consistent with the classification results, larger YOLO variants tended to overfit, whereas medium-sized models demonstrated the best balance between precision and generalisation.
Overall, the findings suggest that classification models are well-suited for controlled experimental settings with individual plant images, while object detection models provide greater applicability in field environments where simultaneous localisation and stress classification are required. The study underscores the importance of aligning model complexity with dataset characteristics and highlights the potential of deep learning for scalable, automated plant stress monitoring in precision agriculture.
