Geographia Technica, Vol 21(2), Special Issue: Artificial Intelligence Applications in Geography, 2026, pp. 206-229
MULTI-ALGORITHM CLASSIFICATION AND CA–MARKOV PREDICTION OF LAND-USE CHANGE ACROSS THREE TEMPORAL INTERVALS: A CASE STUDY OF THE LOE TOR ROYAL PROJECT, TAK PROVINCE, THAILAND
Sorawit ONPHUA 
, Sakhan TEEJUNTUK , Chakrit NA TAKUATHUNG
ABSTRACT: Land-use and land-cover (LULC) change in mountainous watersheds has significant implications for food security, biodiversity conservation, and carbon storage. This study assessed LULC dynamics in the Loe Tor Royal Project area, Tak Province, Thailand, using multi-temporal satellite imagery (2006–2026), five supervised classification algorithms (Random Forest, Artificial Neural Network, Linear Discriminant Analysis, Support Vector Machine, and XGBoost), and CA–Markov spatial modeling across three calibration intervals (3-year, 5-year, and 10-year). Between 2006 and 2026, forest cover declined from 64.10% to 45.16%, while agricultural land expanded from 8.29% to 30.82%, with the rate of agricultural expansion accelerating markedly in recent years. Cramér's V analysis identified distance to settlements (V = 0.85), distance to roads (V = 0.67), and elevation (V = 0.61) as the most influential spatial drivers of LULC transitions. Among the three calibration intervals, the 3-year model achieved the highest validation accuracy, whereas the 10-year model performed substantially worse, indicating that shorter calibration intervals produce more reliable CA–Markov projections. The inferior performance of the 10-year model is attributed to Markov non-stationarity over extended calibration periods and cross-sensor spectral inconsistencies, highlighting the important methodological consideration for CA-Markov applications using multi-sensor datasets. Under Business-as-usual scenarios, forest cover is projected to continue declining, with agricultural land expanding progressively along low elevation zones near road networks and settlements, reaching approximately 42.7% by 2035 under the 3-year scenario. These findings highlight the ecological vulnerability of this critical watershed and emphasize the need for protective buffer zones, conservation agriculture practices, and degraded land rehabilitation efforts.
Keywords: LULC change; CA–Markov model; Multi-algorithm classification; Calibration interval comparison; Sentinel-2; Landsat; Watershed management.
Full article here