Effects of machine learning models and spatial resolution on land cover classification accuracy in Dali County, Shaanxi, China

Land use and land cover (LULC) has undergone drastic changes with the rapid growth of the global population, economic development, and the expansion of agricultural activities. However, the uncertainty of classification algorithms and image resolution based on satellite data for land cover mapping, particularly cropland cover mapping, needs to be investigated sufficiently. In this study, the influence of different spatial-resolution images on classification results was explored by comparing the differences between four machine learning algorithms for LULC mapping. The classification results of this model were also compared with existing global land cover datasets to determine whether the model was capable of producing reliable results. According to the results of this study, the random forest (RF) classifier outperformed the support vector machine (SVM), decision tree (DT), and artificial neural network (ANN) with an overall accuracy (OA) and kappa coefficient of 81.99% and 0.78, respectively. However, SVM and ANN showed greater accuracy on the water class and unused land class, respectively. With increasing spatial resolution, RF’s accuracy increased initially and then decreased when classifying images with five different spatial resolutions (30 m, 16 m, 10 m, 8 m, and 2 m). In particular, with an OA of 82.54% and a kappa coefficient of 0.78, RF performed the best on images with 8 m resolution. Additionally, the RF-based image with 8 m resolution produced a higher OA of 0.88 for cropland. Topography is the main factor that determines the classification performance of different-resolution images. The classification accuracies of RF10 m and RF30 m (10 m and 30 m resolution images, respectively, using RF) were higher (OAs of 93.59% and 94.59%, respectively) than those of the global land cover dataset (LC10 m and LC30 m, land cover images with 10 m and 30 m resolution, respectively), whose high-resolution images showed more details of the land cover. The results of this study highlight that classification algorithms and image resolution are the sources of uncertainty for land mapping. Obtaining reliable land cover mapping requires the use of appropriate classification algorithms and spatial resolution. With these results, it will be possible to develop a national land monitoring system and basic ecological climate models using LULC.