Page 5 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

There is errors and uncertainty in LSE estimation methods, due to its non-uniformity and changes through vegetation
and physical parameters such as texture, composition, surface moisture, roughness, and view angle [14]. In particular, the errors
of LSE and consequently the LST retrieval [13]. Moreover, a few studies have shown that the errors and uncertainty in LSE
may result in an error of 1 K to 2 K in LST using single channel (SC) algorithm around 10 microns [3]. Thus, a small improvement
in LSE can influence LST remarkably. Hence, accurate and robust estimation of LSE remains a challenging problem because of
the uncertainties by residual atmospheric effects, insufficient correction of sun-target-satellite geometry, and uncertainties in
LSE [15]. In particular, LSE in heterogeneous areas such as soil types, rock, urban areas, plant and soil composition shows more
errors and uncertainties [16]. To overcome the errors of the LSE estimation a big project was conducted by National Aeronautics
and Space Administration’s (NASA) Jet Propulsion Laboratory (JPL), California Institute of Technology. To accomplish such
important task, JPL use 15 years data from all available cloud free ASTER product (the results from temperature/ emissivity
separation (TES) algorithm) during 2000-2015 and developed several high spatial resolution mean emissivity database - termed
the ASTER Global Emissivity Datasets (ASTER-GEDv2, GEDv3 and GEDv4) [17, 18]. Likewise, the University of Wisconsin
Global Infrared Land Surface Emissivity Database (UWIREMIS) developed a monthly emissivity data set derived from
emissivity retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS) on the NASA Aqua satellite [19].
Indeed, these big projects are time and cost consuming to obtain emissivity and to overcome the problem of uncertainty as much
as possible. One of the main approaches for controlling errors and uncertainty is using ensemble methods. Evidently, the
ensemble classifier is an appropriate approach to overcoming uncertainties in classification tasks. Advantages of utilizing
ensemble methods on LSE estimation methods are; i) they try to exploit the local different behavior of the individual LSE
methods to improve the accuracy of the overall results. ii) ensemble methods can diminish or eliminate uncertainties in the LSE
estimation, because the LSE value at each pixel adapted from one of the individual LSE methods that has greatest impact. We
were inspired by this idea to use knowledge based algorithms to overcome the errors and uncertainties in the LSE estimation
methods.
Therefore, in this study, a knowledge based approach for estimating LSE is proposed to overcome the errors and
uncertainties in the LSE estimation and LST retrieval. This knowledge based method (KBM) require a global scale information
source of emissivity. The proposed Method receive the obtained results of several individual methods of LSE as input. Moreover,
in KBM, the LSEs are combined based on the error of individual methods of LSE obtained through a knowledge source of LSE.
Indeed, the proposed method encompass the benefits of all individual methods and tackle their shortcoming to not only increase
the accuracy of LSEs but also improve the consistency of LSEs for the whole image pixels. The effectiveness of proposed method
is empirically tested over real data sets. Having LSEs, the LST can be retrieval by different methods. There are many LST
retrieval methods for Landsat data in literature [20-24], as it has a relatively long data record period, since the launch of the first
Landsat [25-27]. In contrast to previous Landsat satellites, the Thermal Infrared Sensor (TIRS) of LDCM data contains two
thermal channels, which split-window (SW) and SC algorithms are capable to utilize for LST retrieval[28]. Considering the strip
problem, ghost signal caused by stray light and a time-varying absolute calibration error for TIRS, the validation exercise is still
a tough problem [29]. Therefore, based on the USGS recommendation on the LDCM data, the single-channel (SC) algorithm of
[30]is used only band 10. To evaluate the impact of the LSE improvement on LST, we proposed an alternative scaling method
based on LST products of MODIS for LST cross-comparison. This paper is organized as follows: In the introduction section, the
problems of LSE estimation, the need to develop an approach to lessen the errors and uncertainty of LSE, a brief description of
the most common methods as well as the objectives of this research are presented. In the next section, the proposed method is
explained in detail. In section 3, the results of the methods and experiments are analysed and finally in section 4, conclusions are
given.

2. The proposed method
Indeed, difficulties and problems in the retrieval of LST from satellite TIR data have two atmospheric and emissivity
effect terms [3]. For a given temperature, when the total water vapor (WV) content in the atmosphere is low, the drier atmosphere,
the impact of the LSE on LST retrieval is large [13] . In addition, in areas where the LST is high, the LSE estimate has the
greatest impact on LST. These two characteristics (i.e. drier atmosphere and the high LST) are features of our study area.
Evidently, various LSE estimation methods yield slightly distinct results for different pixels and classes. Indeed, most of the LSE
estimation methods suffer from uncertainty because of changes in physical and environmental parameters (e.g. soil moisture,
texture, composition, roughness and mixed pixels) and consequently spatial variability of pixel’s emissivity even in a given class.
On the other hand, depending on the image scene conditions, each LSE estimation method has its own capabilities. Let us assume
that we have n LSE estimation methods and each individual method estimates the emissivity based on its assumptions and
constraints. Hence, an ensemble of individual LSE methods can be made flexibly based on characteristics of the study area and
sensor data. Now the question arises as to which value of n emissivity values is appropriate for assigning to a given pixel. It is
accepted that there is combine and fusion strategy in n emissivity values. In the KBMs scheme, two methods called Validity
Average (VAvg) and Class Based Validity Average (CBVA) methods are proposed. In KBM methods require an available
knowledge source of emissivity such as in-situ measurement of emissivity or emissivity products of sensors. In the VAvg method,

2