A Modified Deep Learning Approach for Reconstruction of MODIS LST Product

Abstract

This study aims to apply a modified deep learning model to reconstruct cloudy MODIS LST (Land surface Temperature) images. The proposed system was initially designed to colorize a grayscale image with a Convolutional Neural Network (CNN). We modified this approach by training our model using cloudless (clear-sky) MODIS LST data. In the application, 208 cloudless daily MODIS LST images were used. 90% of these images were utilized in the training step, the remaining 10% were used in the testing step. The average RMSE values of each image ranged from 1.76 o C to 4.41 o C. Results proved the significance of the proposed method in the reconstruction of cloudy MODIS LST pixels even with a small dataset.