A wavelet-artificial intelligence fusion approach (WAIFA) for blending Landsat and MODIS surface temperature Vahid Moosavi a, ⁎, Ali Talebi a , Mohammad Hossein Mokhtari b , Seyed Rashid Fallah Shamsi c , Yaghoub Niazi a a Department of Watershed Management Engineering, Faculty of Natural Resources, Yazd University, Iran b Faculty of Natural Resources, Yazd University, Iran c Department of Natural Resources engineering and Environmental Sciences, College of Agriculture, Shiraz University, Iran abstract article info Article history: Received 18 March 2015 Received in revised form 22 June 2015 Accepted 13 August 2015 Available online xxxx Keywords: Artificial intelligence Data fusion Land surface temperature Thermal infrared data Wavelet transform Land surface temperature (LST) is one of the key parameters in the physics of earth surface processes from local to global scales. However, thermal infrared (TIR) images at both high temporal and spatial resolutions are limited because of the technical limitations of current thermal sensors. Therefore, development of fusion models to obtain thermal data in high spatial and temporal resolutions is crucial in environmental studies. This paper presents a hybrid wavelet-artificial intelligence fusion approach (WAIFA)to produce LST data at the spatial reso- lution of Landsat 8 thermal bands. The theoretical basis and the application procedures of the proposed data fu- sion approach are explained. A case study was performed to predict LSTs of six dates in 2014 from March to August in East Azerbaijan Province, Iran. This approach uses powerful non-linear artificial intelligence modeling systems which can cope with the non-linear nature of the land surface temperature data. In addition, multi- spectral bands and different spectral indices are used as well as thermal data in the modeling process to consider the mixture properties of MODIS pixels. Using a 2D wavelet transform to capture the properties of the main sig- nals (original bands) in horizontal, vertical, and diagonal directions to consider the effect of neighboring pixels is the other improvement of this modeling approach. It can also help the model to deal with the non-stationary properties of the satellite and land surface temperature data. The results indicated that the prediction accuracy of the model in different dates varies from 0.47 K to 1.93 K. © 2015 Elsevier Inc. All rights reserved. 1. Introduction Land surface temperature (LST) is one of the main biophysical vari- ables derived from remotely sensed imagery (Anderson et al., 2008). It is widely used for environmental studies such as assessing crop growth (Abuzar, O'Leary, & Fitzgerald, 2009), monitoring vegetation health (Karnieli et al., 2006), agricultural studies (Badeck et al., 2004; Rojas, Vrieling, & Rembold, 2011), estimation of surface energy flux (Muramatsu, Nakayama, & Kaihotsu, 2006), forest burnt area estimation (Ichoku et al., 2003), monitoring evapotranspiration (Anderson, Allen, Morse, & Kustas, 2012; Yang & Wang, 2011) water stress in croplands (Dragutin & Eitzinger, 2007), energy balance (GriendV.D. & Owe, 1993), drought (Karnieli et al., 2010; Sobrino, Gomez, Munoz, & Olioso, 2007) and soil moisture (Hulley, Hook, & Baldridge, 2010). Acquiring satellite images with high temporal and spatial resolutions remains extremely difficult due to satellite technical constrains. It means that thermal imagery with temporal resolution of 1–2 days or less has a moderate to coarse spatial resolution and those with high spa- tial resolution are limited and are with temporal resolution of more than 15 days (Agam, Kustas, Anderson, Li, & Neale, 2007; Jeganathan et al., 2011).For example Landsat produces very useful information (from the aspect of spatial resolution), but the 16-day revisit cycle limited its application in detecting rapid surface changes such as crop-growth, evaporation, and land surface temperature. Therefore, downscaling thermal imagery can be used to cope with this problem. Several studies have been conducted to downscale satellite imagery (Bechtel, Zakšek, & Hoshyaripour, 2012; Bindhu, Narasimhan, & Sudheer, 2013; Dominguez, Kleissl, Luvall, & Rickman, 2011; Inamdar & French, 2009). However, some of them produce TIR data with high spatial but low tem- poral resolution and others generate TIR data with high temporal but usually very coarse spatial resolution. Therefore, development of valu- able techniques such as “Spatial and Temporal Adaptive Reflectance Fu- sion Model (STARFM)” and its derivatives to enhance both spatial and temporal properties of satellite imagery can be of great importance (Amorós-López et al., 2013; Gao, Masek, Schwaller, & Hall, 2006; Gevaert & García-Haro, 2015; Hilker et al., 2009; Zhu, Chen, Gao, Chen, & Masek, 2010). However, these approaches generally used linear models in some parts of their studies. Although linear models have several advantages and have earned their place as the primary tool for process modeling because of their effectiveness and completeness, there are several deficiencies associated with these models (Borel & Gerstl, 1994). Their main drawback is that many real-world phenomena simply do not correspond to the assumptions of a linear model; in these Remote Sensing of Environment 169 (2015) 243–254 ⁎ Corresponding author. E-mail address: moosavi_v66@yahoo.com (V. Moosavi). http://dx.doi.org/10.1016/j.rse.2015.08.015 0034-4257/© 2015 Elsevier Inc. All rights reserved. Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse