Enhancing Early Plant Disease Detection: 1D to 2D Spectral Transformations

Early detection of plant diseases can be challenging due to the subtle differences in early-stage symptoms. Relying solely on the visible traits is insufficient for accurate diagnosis, especially for presymptomatic detection, emphasizing the significance of improved approaches for early disease identification. Spectroscopy serves as a reliable tool in diagnosing asymptomatic diseases, but interpreting raw 1D spectral data is complex due to the indistinguishable patterns among different disease condition. This study presents a novel approach by converting 1D spectral data of cassava leaves into 2D spectral images to reveal the distinguishing features of healthy cassava leaves from those affected with Cassava Mosaic Disease (CMD), and Cassava Brown Streak Disease (CBSD), aiming to enhance classification accuracy. Traditional machine learning techniques such as k-Nearest Neighbors (kNN) and Extremely Gradient Boosting (XGBoost) are compared with deep learning methods, including a 1D Convolutional Neural Network (CNN) based on EfficientNetB0 for direct spectral classification and several pretrained CNN models for 2D image data classification, including ResNet50V2, MobileNetV2, EfficientNetB0 and InceptionV3. Different padding strategies for 2D data are evaluated, including no padding, end padding, beginning padding, top-bottom padding, and centered padding. Our findings demonstrate that transforming 1D spectral data into 2D images and applying effective padding techniques can improve classification accuracy by 5.06% compared to 1D data classification. This study signify a promising alternative approach for effective early plant disease detection.