Scholarlink

<p>Reliable forecasting of reference evapotranspiration (ET<sub>0</sub>) is vital for climate-resilient irrigation scheduling, agricultural water resource management, and drought mitigation, particularly in data-scarce regions. This research assesses the effectiveness of three deep learning-based models namely Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and a Projected Layer LSTM (Pro-LSTM) for daily ET<sub>0</sub> forecasting. The ET<sub>0</sub> values were computed using the FAO-56 Penman–Monteith technique with key weather parameters. The models were trained and tested using daily ET<sub>0</sub> data derived from meteorological records collected over 7&#xa0;years (2015–2022) at Barishal, 18&#xa0;years (2004–2022) at Gazipur, and 5&#xa0;years (2015–2020) at Ishurdi. This study's primary contribution is assessing model performance not only on historical test datasets but also for extended forecasting over 730&#xa0;days of unseen future data. On the test dataset, Bi-LSTM achieved the highest accuracy, with correlation coefficients (R) of 0.997, 0.999, and 0.999 at Barishal, Gazipur, and Ishurdi, respectively. For long-horizon forecasting, the Pro-LSTM outperformed others at Barishal, achieving R = 0.967, KGE = 0.966, MAE = 0.100&#xa0;mm/d, and RMSE = 0.201&#xa0;mm/d. At Gazipur and Ishurdi, LSTM performed best with R values of 0.964 and 0.977, respectively for long-horizon forecasting. While Bi-LSTM demonstrated superior performance on the test datasets across all stations, LSTM and Pro-LSTM provided more robust forecasting capabilities beyond the available data. These findings highlight LSTM-based models, particularly LSTM and Pro-LSTM as scalable, low-input, and high-accuracy solutions for ET<sub>0</sub> forecasting. Their integration into smart irrigation scheduling and early drought warning systems holds significant promise for climate-adaptive water resource management. Future research should explore spatial generalization and real-time deployment within decision-support frameworks.</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

Medium-scale projection of reference evapotranspiration beyond available data using sequential deep learning models: a case study from Bangladesh

  • 期刊论文

  • 发表日期：

    2025年11月5日

查看全文
  • Dilip Kumar Roy

摘要

Reliable forecasting of reference evapotranspiration (ET0) is vital for climate-resilient irrigation scheduling, agricultural water resource management, and drought mitigation, particularly in data-scarce regions. This research assesses the effectiveness of three deep learning-based models namely Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and a Projected Layer LSTM (Pro-LSTM) for daily ET0 forecasting. The ET0 values were computed using the FAO-56 Penman–Monteith technique with key weather parameters. The models were trained and tested using daily ET0 data derived from meteorological records collected over 7 years (2015–2022) at Barishal, 18 years (2004–2022) at Gazipur, and 5 years (2015–2020) at Ishurdi. This study's primary contribution is assessing model performance not only on historical test datasets but also for extended forecasting over 730 days of unseen future data. On the test dataset, Bi-LSTM achieved the highest accuracy, with correlation coefficients (R) of 0.997, 0.999, and 0.999 at Barishal, Gazipur, and Ishurdi, respectively. For long-horizon forecasting, the Pro-LSTM outperformed others at Barishal, achieving R = 0.967, KGE = 0.966, MAE = 0.100 mm/d, and RMSE = 0.201 mm/d. At Gazipur and Ishurdi, LSTM performed best with R values of 0.964 and 0.977, respectively for long-horizon forecasting. While Bi-LSTM demonstrated superior performance on the test datasets across all stations, LSTM and Pro-LSTM provided more robust forecasting capabilities beyond the available data. These findings highlight LSTM-based models, particularly LSTM and Pro-LSTM as scalable, low-input, and high-accuracy solutions for ET0 forecasting. Their integration into smart irrigation scheduling and early drought warning systems holds significant promise for climate-adaptive water resource management. Future research should explore spatial generalization and real-time deployment within decision-support frameworks.