Drones and artificial intelligence for studying and mapping rivers

The experiment in Friuli Venezia Giulia

The researchers – joined in the initiative by experts from the Nanjing Hydraulic Research Institute and Hohai University in Nanjing – carried out their work on the River Tagliamento in Friuli-Venezia Giulia, a river considered ‘a truly unique case’. ‘It has remained almost entirely free of human intervention, and for this reason – according to the SNPA, the National Environmental Protection System of which ARPA Valle d’Aosta is also a part – it retains the “natural” behaviour typical of the great Alpine rivers of yesteryear: multiple channels, islands of vegetation and, indeed, a great deal of timber carried by the current’.

Drones and artificial intelligence

Hence the decision to adopt a new approach: drones and artificial intelligence instead of on-foot surveys, which would take a long time and present many difficulties over long stretches. “The study tried a different approach,” the SNPA points out, “flying a drone over the river with a standard camera and training an artificial intelligence system to automatically recognise timber in the images.” The researchers tested various AI approaches, comparing more traditional techniques with two ‘deep learning’ models (the same families of algorithms used, for example, for image recognition in smartphones).

Fewer categories and greater precision

‘Fewer categories and greater precision,’ the SNPA emphasises. ‘Asking the system to simply distinguish between “wood” and “everything else” worked better than asking it to recognise water, stones and soil separately. Simplifying the task, in this case, has improved reliability.” This is not an end in itself but a starting point, because, as the experts point out, the most important aspect, from a practical point of view, “concerns the model’s ability to ‘generalise’: the researchers trained it on a stretch of river at a specific time, then tested it across a wider area and on images taken at a different time of year.”

A replicable study

‘The result was positive: the system continued to recognise the timber with a good degree of reliability even under new conditions, albeit with a slightly higher margin of error during the driest months, when dry wood is more easily confused with the surrounding gravel.’

The result is that the tool is cost-effective and can be replicated. ‘A drone fitted with a standard camera,’ conclude the SNPA, ‘combined with artificial intelligence, can help those responsible for river management to understand more quickly and effectively how rivers change over time, with direct implications for the protection of river ecosystems and the planning of land-use management measures. Fewer manual surveys, more data, and monitoring tools that are more accessible even for large areas that are difficult to cover on foot.’