Search

• A proximal sensing platform was built and evaluated for in-field hyperspectral imaging. • 750 nm, 550 nm and 680 nm were most discriminating for Alternaria solani detection. • Pixel based accuracy was 0.92, object based precision was 0.22 and recall 0.84. Automatic detection of early blight caused by Alternaria solani could promote a drastic reduction in the consumption of plant protection agents and the related production losses. A proximal sensing platform was constructed and calibrated for acquiring high resolution hyperspectral images in the field, and used to accurately map Alternaria lesions. High resolution canopy reflectance images were obtained for 32 potato plants that had been infected with A. solani and 32 healthy reference plants. Spectral classifiers like partial least squares discriminant analysis (PLS-DA) and support vector machines (SVM) based on PCA scores were tested to discriminate affected and non-affected pixels. Both spectral classifiers performed well at pixel level with accuracies above 0.92. The NIR region (750 nm) was identified as the most discriminative part of the spectrum for detecting the lesions. As the disease pressure is typically expressed as the number of lesions per area, the accuracy was also evaluated at this level. This indicated a considerable number of false detections at the edges of the leaves and the leaf axils. Therefore, a decision tree was designed based on expert knowledge about the shape of Alternaria lesions, and used to post-process the classified images. This reduced the number of false detections, increasing the precision from 0.17 to 0.22 at the expense of a reduction in recall from 0.88 to 0.84. This leaves considerable room for improvement in the classification accuracy at the object level. We learned that (1) few, broad wavelengths are sufficient and (2) spatial context is essential for the detection of lesions caused by Alternaria infection. The application of more powerful object classification techniques such as convolutional neural networks to enhance the model performance by efficiently encapsulating the spatial context in the classifier might further improve the detection performance. This could pave the way to UAV or tractor based Alternaria mapping. [ABSTRACT FROM AUTHOR]