Robotic Detection of Disease Stress Using Hyperspectral Camera

Automatic early detection of anomalies in plant’s status is crucial for advanced crops management and it became an important topic of multidisciplinary research in the last two decades. More recently, focused research has been addressed to plant’s diseases detection, since the capability of detecting initial infection areas will enable to take targeted correction actions, permitting the potential reduction of pesticides or other input in agricultural systems, a more accurate control of condition of the crop, and eventually a final reduction in inputs costs. In this work we explored the capabilities of automatic detection of disease symptoms in cucumber plants by means of a high resolution hyperspectral camera (400 nm – 1000 nm) an integrated on a 6-DOF robotic manipulator. Thanks to the precise motion control of the hyperspectral camera, the system allows to acquire a spectral hypercube, i.e. a 3D array containing VIS-NIR spectra (nominal resolution of 0.6 nm) of 2D imaged areas scanned while shifting the camera frontally to the target at a speed suitable to obtain the desired spatial resolution (in our case 0.3 mm/pix). For the integration of the camera on the manipulator, a custom end-effector was designed, and a software tool was developed with Matlab to control the manipulator kinematics, the focusing of the camera and the acquisition (scan) of hypercubes. General aim of the implemented system was to obtain repeatable conditions during hypercube acquisitions in order to follow the disease symptoms evolution in the canopy, and to allow controlled multiple acquisitions from different view angles to the plant. The study was conducted on cucumber plants exhibiting Powdery mildew fungal disease symptoms. From the hypercubes, a number of ROI corresponding to helathy and diesead portion of leaves were manually extracted with a custom software tool. The sub-hypercubes were processed at pixel level by applying Principal Component Analysis, Linear Discriminant Analysis, and a combinatorial selection of the most significant wavelengths in discriminating healthy tissue and symptoms. An additional approach based on a morphological analysis of the whole image at the selected wavelength was integrated to the spectral based classification. In particular, the morphological analysis was based on the evaluation of the variance of the spectral intensity in a fixed-sized window moving along the entire image. Very promising results were obtained in terms of recognition of infected portions of leaves and for human machine interaction. As a preliminary results of the study, an average correct discrimination of 89% of pixels of the diseased areas was obtained. Furthermore the integration of the sensor and the manipulator control can be considered completed and ready for a full scale experiment.