# Importance of Flow Metrics on Modeling Macroinvertebrate Community in Dammed Rivers: An Approach With Optimized Gradient Boosting

## TL;DR

This study uses optimized machine learning models to better predict how dams affect river habitats and macroinvertebrate communities.

## Contribution

The study introduces dam impact metrics into habitat modeling and shows that optimized gradient boosting algorithms improve prediction accuracy.

## Key findings

- Gradient boosting algorithms (XGBoost and LightGBM) with optimized parameters showed the highest accuracy in predicting macroinvertebrate habitats.
- Incorporating dam metrics and flow alteration data significantly improved model accuracy for macroinvertebrates.
- Clingers were particularly responsive to low-flow metrics caused by dams.

## Abstract

Habitat models that can predict the habitat suitability for riverine organisms and their distribution along environmental gradients are helpful in watershed environmental management. However, the impacts of dams on riverine communities and their habitats have not yet been considered in such models, although dams have considerably altered important riverine habitats. In this study, we aim to develop catchment‐scale habitat models of macroinvertebrate communities with predictor variables characterizing the impacts of dams. We studied the Omaru River catchment in southwest Japan, where the river flow has been altered considerably because of multiple hydropower dams. Multiple machine learning techniques, such as XGBoost and LightGBM were used to model the habitat distributions of 170 macroinvertebrate taxa observed throughout the river catchment. We used predictor variables of dam impacts derived based on geographical information system data (hereafter, dam metrics) and physically simulated flow data using a hydrological model. Among the modeling techniques, gradient boosting algorithms (XGBoost and LightGBM) with optimized tree number parameters exhibited the highest mean accuracy among the analyzed taxa, followed by the random forest algorithm. The accuracy of the habitat models for the macroinvertebrate community and habitat groups considerably improved with the integration of dam metrics and flow predictors. Of the habit groups considered, clingers showed a keen response to low‐flow metrics, presumably owing to flow alteration caused by the studied dams, as the downstream sections of these dams received only residual flow. Our findings indicate that (1) variables of dam impacts greatly improve the predictive capability of macroinvertebrates and (2) gradient boosting machines with optimized parameters are favorable for habitat modeling of the biotic community. Our models are helpful when river practitioners implement conservation measures as they better understand the environmental consequences of their flood protection designs.

This study developed catchment‐scale habitat models of macroinvertebrate community with predictor variables that characterize dam impacts based on either simple geographical information system data or physically simulated flow data using a distributed hydrological model. Among the modeling techniques, XGBoost and LightGBM with optimized tree number parameters showed the highest mean accuracies among the taxa, followed by random forest. The accuracy of each algorithm greatly improved with the dam metrics and flow alteration metrics in many macroinvertebrates.