Sub-pixel land cover classification in a resource constrained environment: One study area, three algorithms and seven images - what can we learn?

Abstract

Statistical approaches can help environmental managers in resource constrained environments to better manage the areas for which they are responsible by enabling improved decision making through the use of sub-pixel classifications of easier, cheaper and more frequently available Sentinel-2 data. For these managers it is important to know if the same method would yield the best classifications for all seasons and years for their study area. It is known from literature that for different studies, the best algorithm varies, however, this study considers if the same holds true for the same study area across different seasons and years. Seven Sentinel-2 images across two seasons and four years are classified using three algorithms which are both seasonally and own date trained. Results show that the wet season imagery classifications are more accurate than those from the dry season for both seasonal and own date classifications. It is also shown that seasonal classifications are more acceptable for the wet season classifications, while for the dry season there are larger differences between the accuracies for seasonally and own date trained classifications. Finally, it is determined that even for the same study area and season, the best algorithm varies for different images.