Comparative evaluation of Random Forest, SVM, and Naive Bayes algorithms for land cover classification using Sentinel-2 data on Google Earth Engine: A case study in Thai Nguyen, Vietnam

Abstract

This study evaluates the performance of three machine learning algorithms, namely 
Random Forest (RF), Support Vector Machine (SVM), and Naive Bayes (NB), for land cover 
classification using Sentinel-2 satellite imagery in Thai Nguyen province, Vietnam. The imagery 
was processed on the Google Earth Engine (GEE) platform, integrating spectral bands, spectral 
indices, topographic variables, and image texture features to construct the input dataset for the 
classification models. A total of 18,524 sample pixels were used, of which 70% were allocated for 
model training and 30% for accuracy assessment. The results indicate a clear difference in 
classification performance among the algorithms. RF achieved the highest accuracy with an 
overall accuracy (OA) of 90.27% and a Kappa coefficient of 0.880. SVM also produced good 
classification results, with an OA of 88.78% and a Kappa value of 0.862. In contrast, NB showed 
significantly lower performance, with an OA of 37.41% and a Kappa coefficient of 0.238. These 
findings suggest that RF and SVM are capable of effectively modeling nonlinear relationships 
between spectral variables and land cover classes, whereas the independence assumption of NB 
is not well suited to the complex spectral characteristics of multispectral remote sensing data. 
The study highlights the potential of integrating Sentinel-2 data with machine learning algorithms 
on the GEE platform for land cover classification and mapping, and provides a scientific basis 
for selecting appropriate classification algorithms in remote sensing studies.