Sampling approaches for one-pass land-use/land-cover change mapping

In land-use/land-cover change (LULCC) mapping, training fields for changed classes are often difficult to identify. In this study, a new sampling strategy is proposed for mapping LULCCs, in which change and no-change samples are obtained directly from overlaid bi-temporal images with a small shift or rotation. In this way, the artificial changes created maintain a certain degree of geographic information. This method is compared with a simulated sampling approach in which training samples of each land-use/land-cover type are selected separately from individual images and then cross-combined to form the ‘from–to’ classes. The case study demonstrates that both sampling strategies ease the difficulties in performing one-pass classification for LULCC detection and yield more accurate LULCC maps than that of the traditional two-step post-classification comparison method. Between the two sampling strategies, the proposed one provides higher testing accuracy than the simulated sampling approach, while the latter is easier to implement.