AlgarroboNet: An efficient model for the identification of Algarrobo plus trees using aerial images and convolutional neural networks

The identification of seed trees, or 'plus trees,' is key to conserving endangered species such as the Algarrobo (Neltuma pallida). Unmanned aerial vehicles, combined with convolutional neural networks (CNNs), offer an efficient solution. However, the increasing complexity of these networks poses the challenge of balancing performance with computational resources. This study compared the impact of two CNNs, GoogleNet and a network called AlgarroboNet, in the classification of plus trees. Using aerial images from a dry forest in Peru, both networks were trained 30 times and evaluated for accuracy and F₂-measure. GoogleNet has 8 times more layers and 4.2 times more trainable parameters than AlgarroboNet, but only outperformed it by 4 %. The study concludes that it is crucial to adjust model complexity according to the specific task, avoiding unnecessary over-sizing. Future research should focus on adapting models to the specific nature of the task, rather than indiscriminately increasing complexity.