Scholarlink

<p>Accurate estimation of land surface temperature (LST) is essential for environmental monitoring and management applications. While recent advances in remote sensing have made the retrieval of LST from satellite images a common practice in Malaysia, the frequent existence of cloud cover throughout the year makes the retrieval challenging and results in a significant amount of missing LST data. In this study, multiple machine learning models, support vector regressor (SVR), multilayer perceptron (MLP) and random forest (RF) and deep learning models, long short-term memory (LSTM) and gated recurrent unit (GRU), were used to estimate the daytime LST based on significant meteorological and remote sensing variables selected using feature selection methods. This study was performed at two stations, Alor Setar and KLIA Sepang station, which are situated across the Peninsular Malaysia. The meteorological data were sourced from the Malaysian Meteorological Department (MMD), while the MODIS/Terra remote sensing data were retrieved from the Google Earth Engine (GEE) platform. Most of the variables demonstrated moderate to strong correlations with the daytime LST. Maximum air temperature (T<sub>max</sub>) was found to be the most critical variable that cannot be ignored at both study stations. The significance of near infrared (NIR) and shortwave infrared bands (SWIR7) indicated that the daytime LST was strongly influenced by the differences in moisture content. All the models showed satisfactory capabilities in estimating the daytime LST. The overall coefficient of determination, R<sup>2</sup> and Kling–Gupta efficiency (KGE) obtained across the stations ranged from 0.644 to 0.794 and 0.584 to 0.873, respectively. For prediction errors, the values ranged from 1.237℃ to 1.656℃ for mean absolute error (MAE), 0.036 to 0.049 for normalized mean absolute error (NMAE), and 1.605℃ to 2.133℃ for root mean square error (RMSE). Comparing among the models, SVR outperformed the other models in the majority of the scenarios, followed by GRU, LSTM, MLP and RF. The present study has helped in expanding the available predictor variable space and confirmed the feasibility of using AI-based models for LST estimation in Peninsular Malaysia, supported by satisfactory predictive performance metrics reflected in low error values and favourable agreement between observed and predicted values.</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

Integrating remote sensing and meteorological data for AI-based land surface temperature prediction with feature selection approaches

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月24日

查看全文
  • Tze Ying Fong,
  • Yuk Feng Huang,
  • Ren Jie Chin,
  • Chai Hoon Koo

摘要

Accurate estimation of land surface temperature (LST) is essential for environmental monitoring and management applications. While recent advances in remote sensing have made the retrieval of LST from satellite images a common practice in Malaysia, the frequent existence of cloud cover throughout the year makes the retrieval challenging and results in a significant amount of missing LST data. In this study, multiple machine learning models, support vector regressor (SVR), multilayer perceptron (MLP) and random forest (RF) and deep learning models, long short-term memory (LSTM) and gated recurrent unit (GRU), were used to estimate the daytime LST based on significant meteorological and remote sensing variables selected using feature selection methods. This study was performed at two stations, Alor Setar and KLIA Sepang station, which are situated across the Peninsular Malaysia. The meteorological data were sourced from the Malaysian Meteorological Department (MMD), while the MODIS/Terra remote sensing data were retrieved from the Google Earth Engine (GEE) platform. Most of the variables demonstrated moderate to strong correlations with the daytime LST. Maximum air temperature (Tmax) was found to be the most critical variable that cannot be ignored at both study stations. The significance of near infrared (NIR) and shortwave infrared bands (SWIR7) indicated that the daytime LST was strongly influenced by the differences in moisture content. All the models showed satisfactory capabilities in estimating the daytime LST. The overall coefficient of determination, R2 and Kling–Gupta efficiency (KGE) obtained across the stations ranged from 0.644 to 0.794 and 0.584 to 0.873, respectively. For prediction errors, the values ranged from 1.237℃ to 1.656℃ for mean absolute error (MAE), 0.036 to 0.049 for normalized mean absolute error (NMAE), and 1.605℃ to 2.133℃ for root mean square error (RMSE). Comparing among the models, SVR outperformed the other models in the majority of the scenarios, followed by GRU, LSTM, MLP and RF. The present study has helped in expanding the available predictor variable space and confirmed the feasibility of using AI-based models for LST estimation in Peninsular Malaysia, supported by satisfactory predictive performance metrics reflected in low error values and favourable agreement between observed and predicted values.