Dragonfruit Stem Health Classification with Deep Learing and Attention Mechanisms

Abstract

Plant disease detection is important for maintaining the health and quality of crop yields. For the detection of health problems in images of dragon-fruit (Hylocereus) stems, we present a deep learning architecture that is aug-mented by the application of spatial, channel, and domain-specific attention mechanisms. For the purpose of improving accuracy and robustness, the model architecture is combined using features from ResNet50 and EfficientNetB0 back-bones, along with separate attention branches. Training and testing were per-formed using a proprietary dataset of images of dragonfruit stems that are healthy as well as showing various levels of disease. The model initially used frozen fea-ture extractors for training, which were subsequently fine-tuned to improve over-all performance. Experimental results provide high accuracy for classification, with ROC-AUC values above 0.94 for all classes. The proposed method enables precision agriculture operations by providing a robust method for early detection of dragonfruit diseases. Future work will include adapting the model to real-world field images and further optimising it for application in agricultural moni-toring systems.