Land use/land cover classification and scale effect analysis for a multi-temporal superpixel-based segmentation using PlanetScope data

Land-use land-cover (LULC) classification has long been an important topic in Earth observation research, frequently evaluated with recent advances in remote sensing science. This study evaluated the accuracy and suitability of LULC classifications based on the scale effect of a multi-temporal superpixel-based segmentation using PlanetScope (PS) data. We applied the Simple Non-Iterative Clustering (SNIC) algorithm testing five scale factors: 20, 50, 80, 110, and 140. We extracted statistical information of PS bands and vegetation indices from image-objects as input information for classification. In addition, segmentation tests were evaluated by analyzing the variability inside image-objects. Our results showed that the scale factor of 50 presented the highest accuracy while the scale factor of 20 returned the poorest. The scale factor of 20 also created a large number of image-objects inside land parcels, while scale factors of 110 and 140 merged adjacent areas. Segmentation evaluation demonstrated that a satisfactory scale factor for classification is essential once it directly affects the within-class variability and spoils segmentation suitability. The evaluation of these classifications has provided important insights into the effect of the scale factor in high-resolution imagery.