EXPRESS: Early Detection of Huanglongbing in Citrus Trees using Raw Raman Spectra and Deep Residual Networks

Huanglongbing (HLB) is a devastating citrus disease that requires rapid, non-invasive detection methods, especially during its asymptomatic phase. In this study is evaluated the performance of machine learning models and with 1D ResNet-50 and ResNet-152 models were trained from scratch and via transfer learning, for HLB detection using Raman spectroscopy by comparing raw versus pre-processed spectral data to classify Raman spectra from control, symptomatic, and asymptomatic citrus leaves employing raw data and pre-processed spectral data.The ResNet-152 architecture, which was trained from scratch on raw spectra, exhibited superior performance in comparison to all other models, achieving a value of 0.836 ± 0.025 in the balanced accuracy metric and an area under the ROC curve of 0.945 ± 0.011. Additionally, it attained values of 0.848 ± 0.034 for precision and 0.836 ± 0.025 for recall. It is of particular significance that raw data, incorporating biological fluorescence, yielded superior diagnostic markers in comparison to pre-processed spectra, thereby significantly enhancing the identification of asymptomatic samples.In the context of machine learning models, Random Forest demonstrated optimal performance, attaining a balanced accuracy of 0.842 ± 0.029, an F1-measure of 0.852 ± 0.028, an area under the ROC curve of 0.957 ± 0.003, and precision and recall values of 0.873 ± 0.022 and 0.842 ± 0.029, respectively when is evaluated with raw spectra. It is noteworthy that raw data, incorporating biological fluorescence, yielded superior diagnostic markers in comparison to pre-processed spectra. This finding significantly enhances the identification of asymptomatic samples.Machine learning and deep learning architectures trained on raw Raman signals provide a high-precision, non-invasive tool for early detection of HLB. These architectures are competitive with gold standard and superior to the metrics obtained by an expert in the field. This approach eliminates the need for complex signal pre-processing, facilitating the development of real-time, cost-effective diagnostic platforms for sustainable citrus management.