Modeling plant species segmentation using an advanced U-Net and UAV remote sensing: a case study in the High Atlas Mountains of Morocco

Accurately identifying and mapping individual plant species are essential for effective ecological monitoring and conservation of biodiversity. High-resolution Unmanned aerial vehicles (UAV) imagery provides the opportunity to enhance traditional field botanical surveys by capturing ultra-high-resolution spatial details; however, the variability in plant size, shape, and distribution presents significant challenges for precise mapping, even with advanced deep learning methods. The primary objective of this study is to map and identify economically important aromatic plant species in the Timoulit region, located in the High Atlas Mountains of Morocco. To achieve this, we employ UAV imagery combined with an improved U-Net architecture optimized by ResNet50, focusing on species such as Euphorbia abyssinica, Ceratonia siliqua, Juniperus phoenicea, and Pistacia lentiscus. This approach leverages data augmentation and hyperparameter optimization to enhance model generalization and training speed while addressing input constraints. By applying combined loss functions (DiceLoss and CrossEntropy), our model achieves exceptional segmentation performance, with a mean Intersection over Union (MIoU) of 73.99% and a mean pixel accuracy (MPA) of 91.75%. Furthermore, the proposed model outperforms traditional semantic segmentation models, such as DeepLabv3 and MobileNet_v2, demonstrating its efficacy in handling complex plant mapping tasks. Consequently, this method not only improves the accuracy of plant species mapping but also provides a scalable solution for broader environmental applications, thereby supporting sustainable ecosystem management and conservation initiatives.