Deep Classification of Algarrobo Trees in Seasonally Dry Forests of Peru Using Aerial Imagery

Abstract

Seasonally dry forests require tailored strategies to address the challenges posed by deforestation and desertification, particularly for the conservation of protected tree species. Efforts to defend protected flora involve maintaining accurate records of superior specimens of seed-producer trees (plus trees), to support informed management decisions in designated areas. In this paper, a methodology is proposed for the automated classification of plus algarrobo trees (Neltuma pallida), based on RGB aerial imagery and deep learning classifiers. As a proof of concept, three state-of-the-art approaches were evaluated in selected zones of interest, and a new application-specific architecture called AlgarroboNet was proposed for efficient classification of plus algarrobo trees. Initially, a dataset containing the geographic and phenological characteristics of algarrobo trees was provided by the Peru National Forest Service. Subsequent in-situ evaluations retrieved detailed morphometric measurements of plus trees, along with aerial imagery for both plus and non-plus specimens. After correction and segmentation, a balanced database was prepared to evaluate the four deep classifiers. The performance of each approach was summarized using both accuracy and $F_{1}$ -measure, following a hold-out validation strategy with 30 trials. The results reveal that state-of-the-art approaches show a $F_{1}$ -measure that ranges between 0.92 and 0.99 among the models, presenting a trade-off between computational resources and accuracy. Through a detailed analysis, the proposed methodology shows potential for large-scale monitoring and characterization of plus algarrobo trees, and suggest being suitable for implementation, aiming to maintain an up-to-date inventory and enforce reforestation programs.