Almond Tree Variety Identification Based on Bark Photographs Using Deep Learning Approach and Wavelet Transform

Recently, tree species identification has gained in popularity in the silviculture field. As a result, recent works have focussed on recognition of adult evergreen forest trees based on leaves, branching style, bark texture, and, more often, a combination of these features. In this paper, we deal with a new issue: the identification of 2-years-old deciduous fruit-tree varieties within a species, which is a challenging task since we rely exclusively on undistinguishable young bark images. This work is realized to respond to a real need of small farmers interested in planting bare root cultivars. To this purpose, we have built a dataset called AlmBark containing 567 almond bark images with brightness variation and labelled according to the 3 varieties considered (Lauranne, Mazzetto and Independence) which have a high demand in the Tunisian market. We propose a solution combining the Wavelet Transform (WT) technique with a Deep Learning (DL) model for a classification based on almond bark photographs. We investigate the model performances with and without natural background image removal for a mobile implementable solution. Experimentations show that with a background removal step, we obtain an excellent accuracy of \(99\%\) $99\%$ while we obtain a satisfying accuracy of \(96.46\%\) $96.46\%$ using the original photographs.