Modeling Diurnal Land Surface Temperature on a Local Scale of an Arid Environment Using Artificial Neural Network (ANN) and Time Series of Landsat-8 Derived Spectral Indexes


Şekertekin A., Arslan N., Bilgili M.

Journal of Atmospheric and Solar-Terrestrial Physics, vol.206, 2020 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 206
  • Publication Date: 2020
  • Doi Number: 10.1016/j.jastp.2020.105328
  • Journal Name: Journal of Atmospheric and Solar-Terrestrial Physics
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Artic & Antarctic Regions, Communication Abstracts, Compendex, INSPEC, Metadex, Civil Engineering Abstracts
  • Keywords: Land surface temperature (LST), Artificial neural network (ANN), Spectral indexes, SURFRAD, SPLIT-WINDOW ALGORITHM, BUILT-UP, VEGETATION INDEXES, BARENESS INDEX, RETRIEVAL, TM, EXTRACTION, SURFRAD, WATER, CLASSIFICATION
  • Çukurova University Affiliated: Yes

Abstract

This study aims to model diurnal Land Surface Temperature (LST) on a local scale of an arid environment by utilizing the Artificial Neural Network (ANN) and time series analysis of Landsat-8 satellite imageries. An arid region containing an in-situ LST station (DRA) located in Nevada, United States, was chosen as a test site. 78 Landsat-8 satellite imageries covering the test site were utilized to calculate spectral indexes. Since the spectral indexes represent the surface of Earth as land cover indexes, they can be used as indicators that affect the LST. The relationship between ten spectral indexes and in-situ LST were investigated, and the highly correlated indexes were determined as Built-up Area Extraction Index (BAEI) and Normalized Difference Bareness Index (NDBaI). The BAEI and NDBaI showed −0.80 and −0.94 correlation coefficients (r), respectively, with in-situ LST. Those two indexes and meteorological data, namely relative humidity (RH) and air temperature (AT), were used as inputs in the ANN model. A multi-layer perceptron (MLP) feed-forward network was considered in this study. The ANN model presented highly accurate results in the training and testing process with Root Mean Square Error (RMSE) values 0.74 K and 2.54 K, respectively. After learning and testing processes, the weights and biases were extracted to form the mathematical equation of the ANN model, and the equation was utilized to map LST for three data sets which were acquired during the winter and summer times and were not utilized in the ANN model. The results showed that the LST difference was lower than 1 K with regard to wintertime LST images. However, the LST difference was 2.49 K in the summertime. To test the spatial variability of ANN-based LST, MODIS LST products were considered and ANN-based LST was resampled to 1 km, the same resolution as the MODIS data, for comparison. As a result of the comparison, the highest mean LST difference (for all pixels in the scene) between ANN-based and MODIS LST was calculated as −1.1 K. Although the proposed method tended to underestimate LST as it increased, the obtained results showed that the ANN method would be a powerful tool for predicting and modeling the diurnal LST in an arid environment.