Assessing the Effect of Training Sampling Design on the Performance of Machine Learning Classifiers for Land Cover Mapping Using Multi-Temporal Remote Sensing Data and Google Earth Engine

Machine learning classifiers are being increasingly used nowadays for Land Use and Land Cover (LULC) mapping from remote sensing images. However, arriving at the right choice of classifier requires understanding the main factors influencing their performance. The present study investigated firstly the effect of training sampling design on the classification results obtained by Random Forest (RF) classifier and, secondly, it compared its performance with other machine learning classifiers for LULC mapping using multi-temporal satellite remote sensing data and the Google Earth Engine (GEE) platform. We evaluated the impact of three sampling methods, namely Stratified Equal Random Sampling (SRS(Eq)), Stratified Proportional Random Sampling (SRS(Prop)), and Stratified Systematic Sampling (SSS) upon the classification results obtained by the RF trained LULC model. Our results showed that the SRS(Prop) method favors major classes while achieving good overall accuracy. The SRS(Eq) method provides good class-level accuracies, even for minority classes, whereas the SSS method performs well for areas with large intra-class variability. Toward evaluating the performance of machine learning classifiers, RF outperformed Classification and Regression Trees (CART), Support Vector Machine (SVM), and Relevance Vector Machine (RVM) with a >95% confidence level. The performance of CART and SVM classifiers were found to be similar. RVM achieved good classification results with a limited number of training samples.