Application of Long Short-Term Memory neural network model for the reconstruction of MODIS Land Surface Temperature images


ARSLAN N., ŞEKERTEKİN A.

JOURNAL OF ATMOSPHERIC AND SOLAR-TERRESTRIAL PHYSICS, cilt.194, 2019 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 194
  • Basım Tarihi: 2019
  • Doi Numarası: 10.1016/j.jastp.2019.105100
  • Dergi Adı: JOURNAL OF ATMOSPHERIC AND SOLAR-TERRESTRIAL PHYSICS
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Anahtar Kelimeler: Long Short-Term Memory (LSTM), Land Surface Temperature (LST), Recurrent Neural Network, Reconstruction, SPLIT-WINDOW ALGORITHM, GLOBAL SOLAR-RADIATION, EMISSIVITY, ANN
  • Çukurova Üniversitesi Adresli: Evet

Özet

Land surface temperature (LST) is an important parameter that supplies information about the skin temperature of the Earth surface. Remote sensing satellite systems with thermal bands can be used to obtain LST information. One of these satellite systems, namely, Moderate Resolution Imaging Spectroradiometer (MODIS) is mostly utilized in LST studies. One of the problems of obtaining LST from the MODIS data is missing pixels because of the effects such as cloud coverage. This drawback can be encountered by applying Long Short-Term Memory (LSTM) network with one-step-ahead prediction of MODIS data to reconstruct daily LST through the previous data. In this study, LSTM network was applied to the daytime and nighttime MODIS time-series, separately. MODIS LST data (MYD11A1) have the spatial resolution of 1 km x 1 km with 1-day temporal resolution. The selected data range from Day of Year (DOY) 1 in 2017 (01 January 2017) to DOY 59 in 2019 (28 February 2019). MODIS images were processed for the reconstruction of daily LST images concerning an agricultural region in Ceyhan, Adana, Turkey. 82% of data were chosen as the training data while the remaining data were used for testing purposes. The data were reconstructed by feeding the network adding the new data in a moving window in each prediction step. The produced Root Mean Square Error (RMSE) map regarding all reconstruction errors from daytime and nighttime images varied between 2 K to 9 K and 1 K-5 K, respectively. Besides, the coefficients of determination (R-2) at a selected pixel of time-series analysis were obtained as 0.894 and 0.905 for daytime and nighttime LST image, respectively. The results revealed that the LSTM network could be used to fix the missing pixels in LST images.