On Reliability of Remote Sensing Data and Classification Methods for estimating transition rules of the land-use Cellular Automata

Typically, the Cellular Automata (CA) models of Land-Use/Land-Cover (LULC) changes focus on estimating the rules of the LULC changes and analysis of the simulation results. However, the models put aside the uncertainty of the LULC maps that are used for establishing the transition rules. Our study questions the reliability of Remote Sensing (RS) data sources and classification methods applied for constructing these maps. Based on four time intervals within a 36-year period, we construct LULC maps and estimate the transition probabilities between six LULC states. The LULC maps and transition probabilities matrices (TPM) are built based on the manual interpretation of high-resolution aerial photos and classification of multispectral Landsat images for the same years. We consider the maps and TPM derived from the aerial photos as a reference data, and compare them to those constructed from the Landsat images classified with several methods: mean-shift segmentation algorithm followed by Random Forest classification method, and three pixel-based methods of classification: K-means, ISODATA, and maximum likelihood. Then, for each classification the TPM were compared to the referenced TPM. The accuracy assessment of all maps obtained with the pixel-based methods is insufficient for estimating the LULC TPM. The LULC map obtained with the objectbased classification method fit well to that based on the aerial photos, but the estimates of TMP are qualitatively different from those constructed from the aerial photos. This article raises doubts regarding the adequacy of Landsat data and standard classification methods for establishing LULC CA model rules, and calls for the careful reexamination of the entire land-use CA framework.