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