Land use and land cover changes without invalid transitions: A case study in a landslide-affected area

Land use and land cover (LULC) analysis provides valuable information to understand environmental changes and their effects on landslide occurrence. However, LULC time series can be affected by errors in classifications that lead to invalid transitions and, therefore, to misinterpretations. One solution is to include temporal approaches that reduce the effects of invalid transitions. Here, we aimed to evaluate how such methods can improve the LULC analysis for a landslide-affected area. For that, we integrated the Random Forest (RF) class likelihoods with the temporal approach provided by the Compound Maximum a Posteriori (CMAP) algorithm, named here as RF-CMAP. Results from RF-CMAP were compared to those obtained from the traditional RF in a post-classification comparison approach. Although both methods presented high performance, with overall accuracy (OA) values greater than 0.87, RF-CMAP reached higher OA than RF for all the analysed years and corrected 99.92 km(2) (12% of the total area) of invalid transitions presented by the traditional RF. Furthermore, RF-CMAP was capable of correctly classifying more areas than RF in landslides (e.g., 66% and 21% for RF-CMAP and RF in 2000, respectively). Finally, this study contributes to exploring the integration between RF and CMAP algorithms to avoid invalid transitions and to assess how the existence of LULC invalid transitions can impact subsequent analyses.