Performance of ensemble‐learning models for predicting eutrophication in Zhuyi Bay, Three Gorges Reservoir

Abstract

Eutrophication and sporadic algal blooms occurring in the tributary bays of the Three Gorges Reservoir in Hubei, China, have become major environmental issues following impoundment. However, predicting eutrophication with traditional methods based on monthly monitoring data remains challenging. In order to explore the potential of data‐driven models in eutrophication prediction and establish reliable prediction data‐driven model based on monthly monitoring data. In this study, two ensemble‐learning models, random forests (RF) and gradient boosted decision trees (GBDT), were used to predict eutrophication in Zhuyi Bay. To address the target, three objectives were solved. First, RF and GBDT used to regress chlorophyll‐a concentrations showed good model fit across two monitoring data sets, with R2 values of 0.809 and 0.822 for RF and 0.824 and 0.828 for GBDT. Second, the relative variable importance plots computed by ensemble‐learning models was calculated for selecting monitoring parameters and identify drivers of eutrophication. To improve model fit, it was more important to monitor key parameters of eutrophication (such as water transparency) than to increase sample size. Third, K‐Means++ modelling was used to partition eutrophication data into discrete levels. For three eutrophication levels, the classification accuracies of RF and GBDT were 0.8936 and 0.9064, respectively. When using only two eutrophication levels, accuracies for both models increased to 0.9617. This study suggests that ensemble‐learning models, and in particular GBDT (firstly used in eutrophication), show excellent fitting ability for eutrophication compared with other machine‐learning models and provided reliable eutrophication prediction method based on monthly monitoring data.

Improvement in Spatiotemporal Chl-a Data in the South China Sea Using the Random-Forest-Based Geo-Imputation Method and Ocean Dynamics Data

“…The noteworthy performance of both the RF and GBDT models can be attributed to their utilization of ensemble learning methods. These methods leverage multiple decision tree models to generate predictions, resulting in a model characterized by stable performance and superior predictive capabilities [29].…”

“…Gradient Boosted Decision Trees (GBDTs) represent an iterative ensemble learning methodology, orchestrating the sequential construction of an array of decision trees, each poised to incrementally enhance predictive efficacy [29]. At every iteration, the nascent decision tree is engineered with the specific objective of rectifying the residuals stemming from the antecedent model iteration.…”

“…At every iteration, the nascent decision tree is engineered with the specific objective of rectifying the residuals stemming from the antecedent model iteration. Nevertheless, it is imperative to exercise judicious caution when navigating the domain of hyperparameter tuning during the model training process, owing to GBDT's pronounced sensitivity to hyperparameters [28,29].…”

“…The primary objective is to assess the efficacy of several widely employed machine learning algorithms in the task of reconstructing Chl-a concentration patterns specific to this region. The machine learning algorithms employed in this study include Multilayer Perceptron (MLP) [4,27,28], Random Forest (RF) [5,16,25,29,30], Gradient Boosted Decision Tree (GBDT) [28,29], and K-Nearest Neighbor (KNN) [14,18,23,27,31]. Our model differs from previous studies in that it utilizes easily accessible data on ocean dynamical processes as explanatory variables.…”

Advances in regulated river research in China

“…Cascade reservoirs obviously alter instream environmental and hydrological processes. Hu et al (2021) used two ensemble‐learning models (random forests and gradient boosted decision trees) to predict eutrophication based on monthly monitoring data in Zhuyi Bay of the Yangtze River, China. The results illustrate that the performance of gradient boosted decision trees outperformed other machine‐learning models in predicting chlorophyll‐a concentrations.…”