Rehabilitation scenarios for reservoirs: Predicting their effect on invertebrate communities through machine learning

Abstract

Reservoirs are ecosystems that provide important services and reserves of aquatic biodiversity, especially in arid and semiarid regions. However, they are also highly vulnerable to pollution, water abstraction and land use change. Here, we aimed to develop a “dirty‐water” predictive model that could be used to create recovery scenarios and analyze the effect of measures on the biological quality of reservoirs. Accordingly, we constructed a machine learning predictive model that was trained with environmental and macroinvertebrate community data from 129 sites, sampled in six reservoirs in Northeast Brazil. Two different rehabilitation scenarios were simulated (D1 = lower improvement, 25% change; and D2 = higher improvement, 75% change), and three initial levels of disturbance were considered based on PCA analyses: least disturbed, moderately disturbed and severely disturbed. The effects were analyzed in terms of changes in expected taxa (E), observed/expected taxon ratio (OE), biotic and trophic status indices and the spatial distribution of sensitive taxa. The simulations resulted in significant improvements (p < .001) for all disturbance levels and indicators analyzed. Furthermore, rehabilitation measures in artificial systems, such as reservoirs, could result in a higher biodiversity, biological quality and water quality, also in the least disturbed sites. Thus, this approach has the potential to provide information on management and conservation measures, since the simultaneous predictions of many taxa can be used, including: mapping the distribution of species; monitor the population dynamics of species at risk of extinction; and determining the best rehabilitation measures to improve disturbed ecosystems, allowing the determination of the most cost‐effective restoration measures.