Characterization of Hazelnut Trees in Open Field Through High-Resolution UAV-Based Imagery and Vegetation Indices

Abstract

The increasing demand for hazelnut kernels is favoring an upsurge in hazelnut cultivation worldwide, but ongoing climate change threatens this crop, affecting yield decreases and subject to uncontrolled pathogen and parasite attacks. Technical advances in precision agriculture are expected to support farmers to more efficiently control the physio-pathological status of crops. Here, we report a straightforward approach to monitoring hazelnut trees in an open field, using aerial multispectral pictures taken by drones. A dataset of 4112 images, each having 2Mpixel resolution per tree and covering RGB, Red Edge, and near-infrared frequencies, was obtained from 185 hazelnut trees located in two different orchards of the Piedmont region (northern Italy). To increase accuracy, and especially to reduce false negatives, the image of each tree was divided into nine quadrants. For each quadrant, nine different vegetation indices (VIs) were computed, and in parallel, each tree quadrant was tagged as "healthy/unhealthy" by visual inspection. Three supervised binary classification algorithms were used to build models capable of predicting the status of the tree quadrant, using the VIs as predictors. Out of the nine VIs considered, only five (GNDVI, GCI, NDREI, NRI, and GI) were good predictors, while NDVI SAVI, RECI, and TCARI were not. Using them, a model accuracy of about 65%, with 13% false negatives was reached in a way that was rather independent of the algorithms, demonstrating that some VIs allow inferring the physio-pathological condition of these trees. These achievements support the use of drone-captured images for performing a rapid, non-destructive physiological characterization of hazelnut trees. This approach offers a sustainable strategy for supporting farmers in their decision-making process during agricultural practices.

Keywords: Corylus avellana; UAV; aerial photos; non-destructive analyses; phenotyping; vegetation indices.

Figure 1

Aerial images of the two hazelnut fields used in this study: (a) Farigliano and (b) Carrù fields. (c) Representative image of a single hazelnut plant in the Carrù field.

Figure 2

Pre-processing of images for plant recognition and identification of the hazelnut tree canopy. (a) Image collected in the Carrù field showing multiple and overlapping canopies; (b) RGB image collected in the Farigliano field showing well-separated trees; (c) example of application of the normalized difference vegetation index (NDVI) to define the plant contours of the image shown in (b) (NDVI values are shown on the key colors); (d) the same image of (b,c) obtained after excluding pixels with NDVI values < 0.2, with plant contours defined and shown in green.

Figure 3

Example of plant image slicing. The inset represents a magnification of the slice bordered in red. Each slice was visually inspected, and binary classified as healthy/unhealthy.

Figure 4

Distribution of the whole dataset of images collected from hazelnut plants following binary classification in terms of “healthy/unhealthy”, across the whole acquisition period from May to July (three shooting time points).

Figure 5

Boxplots of the vegetation indices calculated on the image dataset.

Figure 6

Performances of the different supervised machine learning algorithms applied to the selected vegetation indices GNDVI, GCI, NDREI, NRI, and GI. The performance is expressed in terms of accuracy and F1-score.