Prediction of MODIS land surface temperature using new hybrid models based on spatial interpolation techniques and...

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Geospatial Analysis

Background:

• Accurate land surface temperature (LST) prediction is vital for climate change, ecology, and environmental studies, requiring models that capture spatial and temporal dynamics.
• Moderate Resolution Imaging Spectroradiometer (MODIS) sensors provide LST data but struggle with missing values due to cloud cover.
• Existing methods for handling missing LST data require improvement for accurate spatiotemporal forecasting.

Purpose of the Study:

• To evaluate the effectiveness of machine learning models, specifically Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Convolutional LSTM (ConvLSTM), for one-step-ahead LST prediction.
• To assess the performance of different imputation techniques, including Inverse Distance Weighting (IDW), Natural Neighbor (NN), and Cubic Spline (C), in addressing missing LST data.
• To identify the optimal combination of machine learning models and imputation methods for accurate spatiotemporal LST prediction.

Main Methods:

• Utilized daily (MYD11A1) and 8-day composite (MYD11A2) MODIS LST data from January 2017 to November 2020 for the Adana province, Turkey.
• Implemented MLP, LSTM, and ConvLSTM models, integrating them with IDW, NN, and C imputation methods to handle missing LST pixels.
• Trained and validated models using 60% of the data, with the remaining 40% allocated for testing (20% validation, 20% test), and evaluated performance using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).

Main Results:

• The combination of ConvLSTM with Natural Neighbor (NN-ConvLSTM) imputation achieved the best performance for daily LST prediction, with an average RMSE of 3.62°C and MAE of 2.85°C.
• The hybrid model combining MLP with Natural Neighbor (NN-MLP) yielded the best results for 8-day composite LST prediction, achieving an average RMSE of 3.57°C and MAE of 2.69°C.
• RMSE maps demonstrated the pixelwise performance, indicating the effectiveness of the proposed hybrid models in spatiotemporal LST prediction.

Conclusions:

• Hybrid machine learning models, particularly NN-ConvLSTM and NN-MLP, show significant potential for accurate one-step-ahead spatiotemporal LST prediction.
• Natural Neighbor imputation proved effective in handling missing MODIS LST data, enhancing the performance of the predictive models.
• The study confirms the capability of integrated machine learning approaches to overcome limitations in satellite-based LST data, providing reliable forecasts for environmental applications.