Water Stress Detection in Potato Crops Using Multispectral Imaging and Advanced Object Detection Models

In recent years, the agricultural sector has been rapidly transformed by the use of advanced technologies, often referred to as 'smart farming' or 'smart agriculture'. Digital tools and other scientific and technological developments have been widely used to revolutionise agricultural practices for better productivity and sustainability. Detecting water stress in plants is one of these challenges to be addressed. Typically, soil moisture sensors are used to assess the condition of the crop. However, the state of the art has identified multispectral imagery as a promising method for detecting water stress in crops, especially using near-infrared (NIR) and red-edge bands. This study investigates the use of multispectral imagery that doesn't only use RGB channels but also NIR and red-edge channels to improve the detection of water stress in potato crops using advanced object detection models. This dataset contains two classes: stressed and healthy, for both RGB and spectral images (Red, Green, Near-Infrared, and Red Edge). There are 300 images used for training and 60 for testing in both the RGB and each spectral band. Additionally, an augmented dataset is provided, consisting of 1500 training images and 60 testing images for both RGB and spectral data. Multispectral imaging, especially incorporating NIR and red-edge bands, offers several advantages over traditional plant sensors, including non-invasive, remote assessment capabilities and the ability to cover large areas more efficiently. These imaging techniques provide valuable insights into plant health by capturing data beyond the visible spectrum, particularly in identifying early signs of water stress. Building on this potential, several studies have been carried out. The first was to investigate the performance of the YOLOv8 model with 4-channel inputs (Red + Green + NIR + Red Edge) on a potato dataset, evaluating the behaviour of different spectral bands in combination with traditional RGB images. The second was conducted to explore the potential of combining multiple channels, specifically RGN (Red, Green, NIR) and RGE (Red, Green, Red Edge), which were used to train a YOLOv8 model that was compared to the same model trained using RGB images. In addition to YOLOv8, a Faster R-CNN model with a ResNet50 backbone was trained and evaluated on the same multispectral image configurations for comparative analysis. Finally, a pre-trained YOLOv8 model from Hugging Face, specifically designed for plant leaf detection and classification, was used for additional investigation and benchmarking. In all of this work, both cross-validation and traditional validation approaches were used to assess model performance. The results of this study show that the combination of multispectral data significantly improves detection accuracy, confirming that the integration of these spectral bands can improve the identification of water-stressed plants compared to RGB-only approaches.