Performance Evaluation of Ensemble Deep Learning Methods for Poplar Plantation Mapping Using Multi-Temporal Sentinel-2 Images
- il y a 1 jour
- 2 min de lecture
Par Yusuf Öztürk (doctorant université de Turquie, accueilli à l’AgroToulouse dans le cadre d'une bourse PRESTJ avec l'ambassade de France en Turquie)
Poplar is among the important species in industrial forestry due to its rapid growth, hybridization capacity, and adaptability to diverse environmental conditions. The mapping of the spatial distribution of poplar plantations and the analysis of their temporal changes are essential for ensuring sustainable resource management and maintaining up-to-date national poplar inventories. Recently, ensemble deep learning approaches, developed by combining the predictions of different deep learning architectures, have attracted increasing attention across various fields, particularly in remote sensing and forestry, due to their high accuracy and strong generalization capabilities in classifying large-scale and multi-temporal data. Within the scope of this study, ensemble deep learning models were developed for mapping poplar plantation areas using multi-temporal Sentinel-2 imagery, through the semantic segmentation-based models under different strategies. A local-to-global mapping approach was adopted for model development and evaluating their generalization potential. In this context, models were trained using multi-temporal Sentinel-2 images acquired from the Akyazi region were used to train models and assess their generalization potential. Then, they were applied to different test areas in Düzce, Terme, and Azizler to assess their spatial generalization performance. Comparative analysis of individual semantic segmentation models revealed that the transformer-based SwinD-Net and Poplar-Net models outperformed traditional models, producing 2–7% higher IoU values across different test areas. On the other hand, the best accuracy values were received by combining a limited number of model predictions rather than integrating all model outputs. The evaluation of results from individual and ensemble deep learning models demonstrated that ensemble approaches provided limited accuracy improvements, remaining below 1% compared to the best-performing individual model. The most robust generalization performance was achieved using ensemble models with heterogeneous structure, incorporating both dataset and model diversity. On the other hand, although ensemble deep learning models yielded limited improvements in accuracy, individual models with more advanced architectures, such as SwinD-Net and Poplar-Net, provided comparable and, in some cases, superior performance. Therefore, considering computational cost, training time, and hardware requirements, single-model designs with deeper architectures represent a more efficient and practical alternative for real-world applications.
