Comparing Supervised and Semi-supervised Machine Learning Methods for Mapping Aquatic Weeds, as Biomass Resource from High-Resolution UAV Images

Detection and mapping are essential steps for resources assessment for biomasses such as aquatic weeds and are also important management steps aimed at curbing the ecological and socio-economic challenges posed by aquatic weeds on waterbodies. However, significant challenges exist in obtaining labeled data to perform supervised classification of aquatic weeds due to risk and cost involved in navigating over large waterbodies. This study assessed whether using semi-supervised learning (SSL) with fewer number of labeled aquatic weeds will be adequate for a machine learning (ML) model to produce high classification accuracy. A simultaneous collection of field samples of the aquatic weed species, mainly water hyacinth and submerged weeds for classification purposes. Image classifications with five ML algorithms (k nearest neighbor — kNN, decision tree — DT, gradient boosting — GB, random forest — RF, support vector machines — SVM) using both supervised and semi-supervised learning methods were performed. The results indicated that both SL and SSL produced high accuracies with F1 score between 0.8 and 0.91 and overall accuracy (OA) between 80 and 95%. It was therefore concluded that semi-supervised classification can equally produce satisfactory accuracy like a supervised classification, though the former has the advantage of using fewer labeled data, thereby reducing cost.