Performance Evaluation of Machine Learning Algorithms in Change Detection and Change Prediction of a Watershed’s Land Use and Land Cover

In recent years, owing to frequent natural and human-caused changes in land use and land cover (LULC), the need to use accurate methods in investigating LULC changes in the study area has gained double importance. In this article, the accuracy and performance of two methods of random forest (RF) and maximum likelihood (ML) were investigated in the classification of the images, determination of factors affecting changes, detection of change, and prediction of LULC. For this purpose, the Gorganrood Watershed was considered the study area. The accuracy assessment results of classified images demonstrated the RF method’s superiority over ML. RF’s kappa coefficient and overall accuracy achieved percentages of (75.31, 79.63%), (78.31, 82.14%), (88.92, 90.9%), and (75.74, 80%), while the ML rates were (54.7, 62.5%), (76.27, 80.35%), (78.79, 82.46%) and (64.69, 70.9%) for the years 1991, 2001, 2011, and 2021, respectively. According to both outputs of the transition potential algorithm, elevation and aspect had the most and least influence on LULC, respectively. Additionally, change detection of classified images using RF and ML revealed that urban fabric, grassland, and permanently irrigated land experienced an increase with percentages of (7.58, 2.52%), (67.75, 60.6%), (15.1, 7.29%). On the contrary, from 1991 to 2021, forest, non-irrigated arable land, and water bodies had reduction rates of (13.42, 12.02%), (75.79, 94.17%), and (56.94, 44.87%), respectively. The results of change prediction using RF also indicated that grassland, forest, non-irrigated arable land and water bodies would decrease by 15.03, 39.65, and 7%, 0.96% during 20 years from 2031 to 2051, whereas urban fabric and permanently irrigated land would increase 17.95, and 44.69%. Similarly, the predicted images using ML depicted that forest and non-irrigated arable land and urban fabric decreased by 14.2, 10.3, and 6.72%, while grassland, permanently irrigated land and water bodies increased by 27.88, 1.54, and 1.18%, respectively. Highlights The evaluation of the accuracy and performance of two machine-learning algorithms in the classification of images, the determination of the influential factors of LULC changes, detection of changes, and prediction of LULC. Using ArcGIS and Terrset software 2020 in the change detection. Using SAGA-GIS 8.0.1 and Envi 5.3 in the classification of images. Using Cramer’s V and P to examine the correlation of variables and LULC categories. Investigating the factors affecting LULC changes, such as slope, aspect, elevation, and distance from the road, distance from urban fabric, and distance from the river using Terrset’s (LCM) section. Predicting future changes using CA-Markov model. Choosing Gorganrood watershed as the study area owing to its diversity from mountains to agricultural land, and evaluating the transition potential.