Convolutional Neural Networks Facilitate River Barrier Detection and Evidence Severe Habitat Fragmentation in the Mekong River Biodiversity Hotspot

Abstract

Construction of river infrastructure, such as dams and weirs, is a global issue for ecosystem protection due to the fragmentation of river habitat and hydrological alteration it causes. Accurate river barrier databases, increasingly used to determine river fragmentation for ecologically sensitive management, are challenging to generate. This is especially so in large, poorly mapped basins where only large dams tend to be recorded. The Mekong is one of the world's most biodiverse river basins but, like many large rivers, impacts on habitat fragmentation from river infrastructure are poorly documented. To demonstrate a solution to this, and enable more sensitive basin management, we generated a whole‐basin barrier database for the Mekong, by training Convolutional Neural Network (CNN)–based object detection models, the best of which was used to identify 10,561 previously unrecorded barriers. Combining manual revision and merged with the existing barrier database, our new barrier database for the Mekong Basin contains 13,054 barriers. Existing databases for the Lower Mekong documented under ∼3% of the barriers recorded by CNN combined with manual checking. The Nam Chi/Nam Mun region, eastern Thailand, is the most fragmented area within the basin, with a median [95% CI] barrier density of 15.53 [0.00–49.30] per 100 km, and Catchment Area‐based Fragmentation Index value, calculated in an upstream direction, of 1,178.67 [0.00–6,418.46], due to the construction of dams and sluice gates. The CNN‐based object detection framework is effective and potentially can transform our ability to identify river barriers across many large river basins and facilitate ecologically‐sensitive management.