Modeling Evapotranspiration using Remotely Sensed Spatial Contextual Information

Abstract

Characterization of the spatial and temporal variabilities of Actual Evapotranspiration (AET) or Latent Heat Flux (λET) from heterogeneous landscapes is essential in studies related to hydrology, climate, agriculture, irrigation, water resources engineering and management and environmental impact assessments. AET/λET is influenced by a large number of factors related to climate, vegetation and soil moisture and therefore its direct measurement is rendered difficult especially over large spatial domains. The only feasible and convenient way to map AET over regional or catchment-scales is through the use of remote sensing technology and accordingly, numerous world-wide studies have focussed on this approach. Among these, the Penman-Monteith (PM) and Priestley-Taylor (PT) methods have proved to be most popular on account being simple but yet providing reasonably accurate estimates of regional AET. However, most previous studies have implemented satellite-based PM and PT AET estimation approaches to crop lands located in arid to semi-arid regions. Therefore, the main focus of the present study is to develop satellite-based AET estimation methods which can be applied to wet tropical regions possessing natural vegetation. The current research work is aimed at the development, application and evaluation of methodologies for estimation of AET/λET by the PM approach using Moderate Resolution Imaging Spectrometer (MODIS) satellite imagery. The feasibility of extracting the bulk surface conductance (Gs); an important parameter in the PM model, from the spatial contextual information present in a scatter plot of Land Surface Temperature (LST) versus Fraction of Vegetation (Fr) is explored in this study. Also, few studies seemed to have compared the performances of the PM and the PT model and other similar models using the same dataset and therefore this exercise was taken up. Using a general expression for Gs derived by assuming a two-source total λET (canopy transpiration plus soil evaporation) approach proposed by previous researchers, minimum and maximum values of Gs for a given region can be inferred from a trapezoidal scatter plot of pixel-wise values of LST and corresponding Fr. Using these as limiting values, Gs values for each pixel can be derived through interpolation and subsequently used with the PM model to estimate λET for each pixel. The proposed methodology was implemented in 5 km x 5 km areas surrounding each of four AsiaFlux tower sites located in different countries of tropical south-east Asia which were selected based on certain specific criteria. MODIS data products of MOD11A1 product of Land Surface Temperature (LST) at 1000 m resolution, MOD09GA product of Land Surface Reflectance (LSR) at 1000 m resolution, MOD15A2 product of Leaf Area Index (LAI) as eight days average composited at 1000 m resolution and Digital Elevation Model (DEM) at 30 m resolution were used. Preliminary processing of images was performed using MODIS reprojection tool (MRT) and was converted to a standard format that can be read by MATLAB software. Geo-referencing, subsetting and pixel-wise analysis corresponding to the study area were performed using ArcGIS and ERDAS IMAGINE. Model evaluation was carried out using the following performance measures- coefficient of determination (R2), root mean squared error (RMSE), percent bias (PBIAS) and the intercept (a) and slope (b) terms of a linear regression fit (y = a + bx). Excellent comparisons were obtained between tower measured λET and those estimated by the proposed approach for all four flux tower locations (R2 = 0.85 – 0.96; RMSE = 18.27 – 33.79 W/m2). The proposed methodology was compared with two alternative methods proposed by previous researchers. Performances of all three approaches were comparable indicating the robustness of the methodology proposed in the present study. The PM method proposed in the present study was implemented in the Hemavathi sub-basin which is located in the Cauvery River Basin, Karnataka, India to map spatial patterns of daily AET. MOD16A2 product of actual evapotranspiration (AET) as eight day average composited at 500 m resolution was used for validation purposes. Climate records for the Belur station were used. The analysis was carried out for two dates in summer and two dates in winter separately for the years 2007 (wet year) and 2012 (dry year). For each date, trapezoidal scatter plots of MODIS-derived LST values versus Fr were plotted by considering 1 km2 pixels in the study area of 304 km2. For each day, estimated AET values by the PT approach (AETPT), PM model of the present study with Ga computed using Leuning equation (AETPM) and with Ga computed using Choudhary equation (AETPMCH) for each of the 304 pixels were extracted and compared with the corresponding pixel-wise MOD16A2 ET estimates. Results of the performance evaluation of AET estimation methods relative to MOD16A2 showed that the PM model proposed in the present study with Ga computed using the Leuning equation (AETPM) performed reasonably well for both the wet and dry years. High values of R2 (0.77 – 0.90) and reasonably low values of RMSE (0.28 – 0.38 mm/day) were obtained but the PBIAS values were somewhat high (-7.04 – -12.41). Also, the PM model yielded relatively poorer estimates for the winter days of the drier year 2012. The performance of the PT model was quite similar to the PM model with similar performance statistics being recorded. However, slightly lower RMSE values were obtained for this model on some days. The PM model proposed in the present study with Ga computed using the Choudhary equation turned out to be the best model as indicated by the lowest values of RMSE (0.19 – 0.25 mm/day), although R2 values were similar. Also, use of the Choudhary equation reduced PBIAS values significantly for all days considered. Using the pixel-wise values of AETPMCH, maps showing the spatial variability were prepared for the Hemavathi sub-basin for all the dates considered in 2007 and 2012. The variations of AET over the basin appear to be influenced by topography, type of LU/LC, LST and moisture availability conditions. The satellite-based spatial contextual information approach adopted in the present study for the first time with the PM model has proved to be a simple, calibration-free and accurate method. As demonstrated by previous studies and also the present study, use of the LST-Fr plot does not require additional hydrological data for optimization of the AET model parameters. The framework for implementing the spatial contextual information approach to derive operational estimates of daily AET over large spatial domains has been developed and validated in this study.