Evidential two-step tree species recognition approach from leaves and bark

Abstract The contribution of this paper is twofold. First, this paper aims at developing an intelligent system that emulates the decision-making ability of a botanist expert in the recognition of tree species from their leaves and bark. The main challenges of this recognition problem are related to the high diversity of trees in nature, the interspecies similarity and the intra-species variability. Therefore, similarities between species cause several confusions during recognition. The proposed decision system is designed to solve this complex problem of tree species recognition by reasoning with knowledge sets where the inference engine is based on belief functions theory, which reduces confusion between species and achieves greater accuracy. Secondly, this paper proposes a practical solution that can be embedded in the user's smartphone without any need for an internet connection. Therefore, our approach is adapted for smartphone limits, i.e. limits related to memory and computation capacity. Once in nature, everybody should appreciate the idea of having a mobile application that reflects the skills and know-how of a botanist. Building an application to make the potential of tree species recognition accessible and easy to use is a challenging problem. From methodological perspectives, the suggested method is a two-step recognition approach that identifies the leaf in a first step and refines the results using the bark in the second step. In fact, the first step is used to reduce the dimensionality of the problem through the identification of a subset of most probable species. The second step is performed using a modified evidential k Nearest Neighbors (EkNN) algorithm that recognizes the bark from the output of the first step. A set of experiments on real-world data is presented in order to study the accuracy of the proposed solution against existing ones.