Development of transferable hybrid deep learning networks for temporal and multi-regional mapping of poplar plantations with Sentinel-2

Abstract

Poplar plantations are vital for short-rotation forestry due to their fast growth, hybrid vigor, and adaptability to diverse environmental conditions. Accurate mapping of these plantations is essential for sustainable resource planning and forest inventory updates. This study introduces three Ensemble Deep Learning (EDL) models that integrate one-dimensional convolutional neural networks (1D-CNN) and long short-term memory (LSTM) networks to classify poplar plantations using Sentinel-2 time series data. These models were tested in Akyazi, Turkiye, and compared with commonly used Ensemble Machine Learning (EML) algorithms, including Random Forest, XGBoost, and CatBoost. Although EML models achieved slightly higher overall accuracy on validation data, significantly lower recall values were observed during temporal testing, with differences reaching up to 6 %, indicating a greater tendency to misclassify poplar pixels across different dates. Among the proposed approaches, EDL-2, designed with a multi-channel CNN-LSTM structure, consistently demonstrated superior accuracy and temporal stability. This architecture effectively captured both spectral and seasonal patterns, enabling reliable classification across time and space. When applied to test sites in Terme and Azizler (Turkiye), F-score values for the poplar class reached 97.2 % and 96 %, respectively, compared to 94.5 % in Akyazi. A four-year change analysis revealed a 2.5 % increase in poplar cover in Akyazi, with declines of 7.5 % and 36 % in Terme and Azizler, respectively. These results highlight EDL-2's strong generalization capacity, validated through successful application to an independent test site in France, demonstrating its potential for large-scale poplar mapping across diverse geographic contexts. (c) 2025 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.