Scholarlink

<p>The potential Evapotranspiration (PET) is a crucial hydrological variable for understanding the water balance and ecosystem dynamics, particularly in different altitudinal Mediterranean forest environments. Traditional empirical methods for estimating PET have frequently failed to generalize accurately to heterogeneous environments, including Mediterranean forests, due to climatic and topographic diversity. This study develops and evaluates bi-directional transferable Deep Learning (DL) models for PET estimation using meteorological and environmental data collected from Mediterranean forest areas, namely Pertouli and Taxiarchis, covering multiple altitudinal levels. The models were trained and evaluated in both directions: first trained on Pertouli and transferred to Taxiarchis, and second trained on Taxiarchis and transferred to Pertouli. These models were deployed and retrained using Transfer Learning (TL) for PET estimation to address the limited data availability at the target site. Several model architectures were implemented and evaluated in different Scenarios. The experimental findings show that the transferable DL models achieve <i>R</i><sup>2</sup> values between 0.7 and 0.87 and RMSE values from 0.55 to 0.75, according to the data availability of the target site and the meteorological data used for training. Among the evaluated model architectures, the RNN model achieved the highest performance with <i>R</i><sup>2</sup> = 0.767, MAE = 0.521, RMSE = 0.668, and sRPIQ = 0.755 when the training data were limited, while the LSTM obtained the highest accuracy under higher data availability conditions, namely, <i>R</i><sup>2</sup> = 0.873, MAE = 0.384, RMSE = 0.55, and sRPIQ = 0.831. This work demonstrates the feasibility of bi-directional model transfer across altitudinal forest gradients under Mediterranean conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transferable deep learning models for the estimation of daily potential evapotranspiration across altitudinal forest gradients in the Mediterranean

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Petros Amanatidis,
  • Stefanos Stefanidis,
  • Konstantinos Ioannou,
  • Nikolaos Proutsos,
  • Ilias Karmiris,
  • Panagiotis Stefanidis,
  • Dimitris Karampatzakis

摘要

The potential Evapotranspiration (PET) is a crucial hydrological variable for understanding the water balance and ecosystem dynamics, particularly in different altitudinal Mediterranean forest environments. Traditional empirical methods for estimating PET have frequently failed to generalize accurately to heterogeneous environments, including Mediterranean forests, due to climatic and topographic diversity. This study develops and evaluates bi-directional transferable Deep Learning (DL) models for PET estimation using meteorological and environmental data collected from Mediterranean forest areas, namely Pertouli and Taxiarchis, covering multiple altitudinal levels. The models were trained and evaluated in both directions: first trained on Pertouli and transferred to Taxiarchis, and second trained on Taxiarchis and transferred to Pertouli. These models were deployed and retrained using Transfer Learning (TL) for PET estimation to address the limited data availability at the target site. Several model architectures were implemented and evaluated in different Scenarios. The experimental findings show that the transferable DL models achieve R2 values between 0.7 and 0.87 and RMSE values from 0.55 to 0.75, according to the data availability of the target site and the meteorological data used for training. Among the evaluated model architectures, the RNN model achieved the highest performance with R2 = 0.767, MAE = 0.521, RMSE = 0.668, and sRPIQ = 0.755 when the training data were limited, while the LSTM obtained the highest accuracy under higher data availability conditions, namely, R2 = 0.873, MAE = 0.384, RMSE = 0.55, and sRPIQ = 0.831. This work demonstrates the feasibility of bi-directional model transfer across altitudinal forest gradients under Mediterranean conditions.