A mixed training sample-based spectral unmixing analysis for improving fractional abundance estimation of Detroit landscape endmembers using Landsat images

The existence of mixed pixels in medium spatial resolution images impedes the accuracy improvement of extracting information of city landscape land use and land cover (LULC) types through image-based classification. Spectral unmixing of images provides the potential for improving accuracy of the information. However, spectral variability of endmembers limits the capacity of traditional pure training sample-based spectral unmixing methods. To overcome this gap, we proposed a novel mixed training sample-based spectral unmixing method in which the spectral reflectance matrix is estimated using randomly drawn training samples instead of subjectively selected so-called pure pixels. The proposed method was validated in Detroit city landscape to estimate fractional abundances of water, tree, urban and grass using 2010 Landsat images and aerial photographs through comparison of four spectral unmixing methods including multiple linear regression (MLR), random forest (RF), artificial neural network (ANN) and convolutional neural network (CNN) with 400 subjectively selected pure training samples, 400 randomly drawn mixed training samples and 400 validation samples. The results showed that 1) The proposed methods significantly outperformed the traditional pure training samplebased methods (p < 5 %) in terms of root mean square error (RMSE), mean residual and absolute mean residual, with reduction of RMSE by 13.4 % for MLR, 14.4 % for RF, 25.6 % for ANN and 29.5 % for CNN; 2) The CNNbased spectral unmixing with the mixed training samples had the most accurate estimates, and the machine learning (ML)-based spectral unmixing methods (RF, ANN and CNN) led to significantly more accurate estimates than the MLR; and 3) With the mixed training samples, there were no statistically significant differences of absolute mean residuals among the ML-based methods. Therefore, this study provides potential to improve estimation of endmember fractional abundances for spectral unmixing of medium spatial resolution images for large city landscapes.