Comparing machine learning techniques for aquatic vegetation classification using Sentinel-2 data

Wetlands, among the most valuable ecosystems, are increasingly threatened by anthropogenic impacts and climate change. Mapping wetland vegetation changes is crucial for conservation, management, and restoration of such sensitive environments. Machine Learning (ML) algorithms, such as Random Forest (RF), Support Vector Machine (SVM), kNearest Neighbour (kNN), and Artificial Neural Networks (ANN) are commonly applied for wetland mapping based on remote sensing data. However, scientific literature on this topic is often biased towards limited study areas, and lacking generalization testing over heterogeneous environmental conditions (e.g. latitude, ecoregion, wetland type). In this study, we compared eight ensemble and standalone ML methods, aiming at finding the best performing ones for aquatic vegetation mapping using Sentinel-2 over nine study areas and different seasons. The classifiers were tested to distinguish nine different classes - five aquatic vegetation classes and four background land cover classes - with seasonal monthly composites (April-November) of spectral indices as input. Results suggest that ensemble methods, such as RF, generally show higher predictive power with respect to most of common standalone classifiers (e.g. kNN or DT), which show the highest level of overall disagreement. SVM method overcame all the other classifiers, both standalone and ensemble, over our reference dataset, scoring an overall accuracy of 0.977 ± 0.001; In particular, SVM was the best over transitional aquatic vegetation classes (helophytes and submerged-floating association), which are the ones most frequently misclassified by other methods. Further developments of this research will focus on assessing the influence on classification performance of predictor variables and variations in input features.