Jump to Main Content
A semi-empirical approach for the estimation of land-surface emissivity from satellite data based on spectral index fusion using ensemble regression
- Jouybari-Moghaddam, Yaser, Saradjian, Mohammad Reza
- International journal of remote sensing 2019 v.40 no.11 pp. 4213-4243
- Advanced Spaceborne Thermal Emission and Reflection Radiometer, Landsat, data collection, land use, models, moderate resolution imaging spectroradiometer, normalized difference vegetation index, reflectance, regression analysis, remote sensing, soil, spatial data, surface temperature, vegetation
- Land Surface Emissivity (LSE) is a key parameter in the thermal remote sensing, with several important applications, most notably in Land Surface Temperature (LST) estimation. This paper presents a semi-empirical method of LSE estimation from remote sensing data based on a fusion of spectral indices using the ensemble regression methods. The performance of the proposed method for Moderate Resolution Imaging Spectroradiometer (MODIS), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS) data was evaluated and compared with other semi-empirical methods developed for these sensors. The proposed method was designed in four stages. In the first stage, the reflectance of non-thermal bands and emissivity of thermal bands were simulated for different classes using the ASTER spectral library and the spectral response function of each sensor. In the second stage, the dataset to be used for the training of ensemble regression was arranged by calculating a number of spectral indices, which constitute the feature space along with non-thermal bands. In the third stage, the regression between emissivity of thermal bands of each sensor and the features extracted in the second stage was derived by the use of bagging, boosting and Random Forest (RF) regression methods. In the final stage Using Normalized Difference Vegetation Index (NDVI) values, the image was categorized into three classes including vegetation, non-vegetation and mixture areas using conditions NDVI > 0.5, NDVI < 0.2 and 0.2 ≤ NDVI ≤ 0.5, respectively. The non-vegetation class was then categorized to soil, rock, and man-made classes using land use map. The spectral indices of these classes were then calculated, and the corresponding model trained in the third stage was used to estimate the LSE for that band. The results of LSE estimations were compared with the standard product of each sensor. Due to the lack of standard product for Landsat-8, the ASTER product was used as a substitute. For better analysis, the proposed method was also evaluated with other semi-empirical methods developed for MODIS, ASTER and OLI/TIRS sensors. This evaluation showed that the lowest Root Mean Square Error (RMSE) values for OLI/TIRS bands 10 and 11 are 0.0070 and 0.0075 obtained, respectively, by bagging and RF regression methods. For ASTER bands 13 and 14, the lowest RMSE values of 0.0078 and 0.0077 are both obtained by RF regression. For MODIS bands 31 and 32, the lowest RMSE values are 0.0053 and 0.0049 and obtained by boosting method. A comparison between the proposed method and other semi-empirical methods provided for these sensors demonstrated the ability of the method to improve the RMSE by up to 0.5%. Regarding the higher accuracy and applicability of the proposed method, it can serve as an effective and efficient means of estimating LSE using remote sensing data.