Detection of a vascular wilt disease in potato (‘Blackleg’) based on UAV hyperspectral imagery: Can structural features from LiDAR or SfM improve plant-wise classification accuracy?

Ensuring plant health is a key factor to maximize crop yield. Despite that, the current field scouting and disease monitoring approaches often rely on visual evaluations and are, therefore, subjective and time demanding. New methods to assist in disease detection and severity assessment are required to allow better crop management and higher throughput in field phenotyping studies. With this objective, techniques involving the use of multi- and hyperspectral imagery for retrieval of plant traits and assessment of general crop health status are increasingly being proposed as alternatives to conventional disease monitoring approaches. Conversely, research focusing on specific pathogens are still lacking in many cases, in particular studies investigating multi-source sensing approaches, which have the potential to improve retrieval/classification accuracy. In this study, hyperspectral imagery and point clouds obtained with LiDAR or through Structure from Motion algorithm (SfM) applied to high resolution RGB images were evaluated as possible alternatives to detect Blackleg (caused by bacteria of the genera Pectobacterium and Dickeya) in potato. It was demonstrated that all the different datasets have potential to discriminate healthy from diseased plants. The combination of Vegetation Indices (VIs) derived from hyperspectral images with structural features from LiDAR resulted in the best validation results (Balanced Accuracy – BA = 0.915). Small improvements were also achieved by combining VIs with SfM features (BA = 0.876) in comparison to VIs alone (BA = 0.846). Evaluation of feature importance for classification models derived from the different datasets indicated that after structural features derived from LiDAR or RGB imagery were added as predictor variables the relative importance of VIs for the predictions decreased, in particular for VIs related to LAI or other traits describing canopy properties. Finally, analysis of false negatives and positives indicated some limitations to the predictive potential of the different datasets, with diseased and healthy plants eventually presenting atypical structural and spectral characteristics in comparison to those expected for their classes. Therefore, multi-source sensing, including additional modalities (e.g., thermal or fluorescence), might be required to further improve detection of pathogens with complex symptoms, as those affecting roots, tubers and stems.