An ensemble pansharpening approach for finer-scale mapping of sugarcane in Landsat 8 imagery

In International Journal of Applied Earth Observation and Geoinformation
Peer-reviewed Article

In this study we tested the impacts of three fast pansharpening methods – Intensity-Hue-Saturation (IHS), Brovey Transform (BT), and Additive Wavelet Transform (AWT) – on the classification of sugarcane in a Landsat 8 image (bands 1-7), and proposed an ensemble pansharpening approach that combines the pixel-level information from the IHS and BT pansharpened images for classification. IHS and BT inject spatial information from the panchromatic band by additive and multiplicative operators, respectively, and it is typically unknown before-hand which is more appropriate for each multispectral band. Support Vector Machine (SVM) is a commonly-used classification algorithm that is relatively insensitive to noisy classification variables if a sufficient number of training samples are provided. Thus an approach that incorporates both IHS and BT for pansharpening and SVM for classification may produce a more accurate classification result. Another benefit of using IHS and BT for this type of ensemble approach is their speed, as the low resolution multispectral image must be pansharpened more than once prior to classification.

To test the proposed ensemble pansharpening approach, the original Landsat multispectral image, the pansharpened images, and the 14-band IHS-BT ensemble image were classified into “sugarcane” and “other” land cover classes using SVM, and the classification accuracy of each image was assessed. Individually, neither IHS nor BT pansharpening led to higher classification accuracy than the original Landsat image, but the ensemble image achieved a higher Kappa coefficient (0.676) than the original Landsat image (0.611) and higher class-specific accuracies. These results indicate that an ensemble pansharpening approach can lead to higher classification accuracy. For comparison, the Kappa coefficient (0.628) and class-specific accuracies of the AWT pansharpened image were higher than those of the original Landsat image, but lower than those of the ensemble approach. Based on the positive performance of the ensemble approach in this study, we recommend further investigation of ensemble pansharpening for image analysis (e.g. classification and regression tasks) in agricultural and non-agricultural environments.