Modelling remote barrier detection to achieve free-flowing river targets

Abstract Fragmentation caused by artificial barriers is one of the main stressors of rivers worldwide. However, many barrier inventories only record large barriers, which underestimates barrier numbers, and hence fragmentation. Corrected barrier numbers can be obtained via river walkovers, but these are costly and time consuming. We assessed the performance of remote sensing as an alternative to river walkovers for barrier discovery by comparing the number and location of barriers detected in the field with those detected using Google Earth imagery. Only 56% of known barriers could be detected remotely, but machine learning models predicted the likelihood of remote detection with 62%–65% accuracy. Barriers located downstream were twice as likely to be detected remotely than those in the headwaters, the probability of detection diminishing by 3%–4% for every decrease in Strahler stream order and for every 10 km increase in distance from the river mouth. Barriers located in forested reaches were 35% less likely to be detected than those in open reaches. Observer skills also affected the ability to locate barriers remotely and detection rate varied by 11% between experienced and less experienced observers, suggesting that training might improve barrier detection. Our findings have implications for estimates of river fragmentation because they show that the most under-represented structures in barrier inventories, i.e. small barriers located in forested headwaters, are unlikely to be detected remotely. Although remote sensing cannot fully replace ‘boots on the ground’ field surveys for filling barrier data gaps, it can reduce the field work necessary to improve barrier inventories and help inform optimal strategies for barrier removal under data-poor scenarios.