Many countries have attempted to mitigate and manage issues related to harmful algal blooms (HABs) by monitoring and predicting their occurrence. The infrequency and duration of HABs occurrence pose the challenge of data imbalance when constructing machine learning models for their prediction. Furthermore, the appropriate selection of input variables is a significant issue because of the complexities between the input and output variables. Therefore, the objective of this study was to improve the predictive performance of HABs using feature selection and data resampling. Data resampling was used to address the imbalance in the minority class data. Two machine learning models were constructed to predict algal alert levels using 10 years of meteorological, hydrodynamic, and water quality data. The improvement in model accuracy due to changes in resampling methods was more noticeable than the improvement in model accuracy due to changes in feature selection methods. Models constructed using combinations of original and synthetic data across all resampling methods demonstrated higher prediction performance for the caution level (L-1) and warning level (L-2) than models constructed using the original data. In particular, the optimal artificial neural network and random forest models constructed using combinations of original and synthetic data showed significantly improved prediction accuracy for L-1 and L-2, representing the transition from normal to bloom formation states in the training and testing steps. The test results of the optimal RF model using the original data indicated prediction accuracies of 98.8% for L0, 50.0% for L1, and 50.0% for L2. In contrast, the optimal random forest model using the Synthetic Minority Oversampling Technique–Edited Nearest Neighbor (ENN) sampling method achieved accuracies of 85.0% for L0, 85.7% for L1, and 100% for L2. Therefore, applying synthetic data can address the imbalance in the observed data and improve the detection performance of machine learning models. Reliable predictions using improved models can support the design of management practices to mitigate HABs in reservoirs and ultimately ensure safe and clean water resources.

Rivers act as natural conduits for the transport of plastic debris from terrestrial sources to marine environments. Accurately quantifying plastic debris in surface waters is essential for comprehensive environmental impact assessments. However, research on the detection of plastic debris in surface waters remains limited, particularly regarding real-time monitoring in natural environments following heavy rainfall events. This study aims to develop a real-time visual recognition model for floating plastic debris detection using deep learning with multi-class classification. A YOLOv8 algorithm was trained using field video data to automatically detect and count four types of floating plastic debris such as common plastics, plastic bottles, plastic film and vinyl, and fragmented plastics. Among the various YOLOv8 algorithms, YOLOv8-nano was selected to evaluate its practical applicability in real-time detection and portability. The results showed that the trained YOLOv8 model achieved an overall F1-score of 0.982 in the validation step and 0.980 in the testing step. Detection performance yielded mAP scores of 0.992 (IoU = 0.5) and 0.714 (IoU = 0.5:0.05:0.95). These findings demonstrate the model’s robust classification and detection capabilities, underscoring its potential for assessing plastic debris discharge and informing effective management strategies. Tracking and counting performance in an unknown video was limited, with only 6 of 32 observed debris items detected at the counting line. Improving tracking labels and refining data collection are recommended to enhance precision for applications in freshwater pollution monitoring.

Chromophoric dissolved organic matter (CDOM) is a mixture of various types of organic matter and a useful parameter for monitoring complex inland surface waters. Remote sensing has been widely utilized to detect CDOM in various studies; however, in many cases, the dataset is relatively imbalanced in a single region. To address these concerns, data were acquired from hyperspectral images, field reflection spectra, and field monitoring data, and the imbalance problem was solved using a synthetic minority oversampling technique (SMOTE). Using the on-site reflectance ratio of the hyperspectral images, the input variables R_rs (452/497), R_rs (497/580), R_rs (497/618), and R_rs (684/618), which had the highest correlation with the CDOM absorption coefficient a_CDOM (355), were extracted. Random forest and light gradient boosting machine algorithms were applied to create a CDOM prediction algorithm via machine learning, and to apply SMOTE, low-concentration and high-concentration datasets of CDOM were distinguished by 5 m⁻¹. The training and testing datasets were distinguished at a 75%:25% ratio at low and high concentrations, and SMOTE was applied to generate synthetic data based on the training dataset, which is a sub-dataset of the original dataset. Datasets using SMOTE resulted in an overall improvement in the algorithmic accuracy of the training and test step. The random forest model was selected as the optimal model for CDOM prediction. In the best-case scenario of the random forest model, the SMOTE algorithm showed superior performance, with testing R², absolute error (MAE), and root mean square error (RMSE) values of 0.838, 0.566, and 0.777 m⁻¹, respectively, compared to the original algorithm’s test values of 0.722, 0.493, and 0.802 m⁻¹. This study is anticipated to resolve imbalance problems using SMOTE when predicting remote sensing-based CDOM. It is expected to produce and implement a machine learning model with improved reliable performance.