Examples from Machine learning algorithm to detect plastic accumulations on beaches and at sea developed by Pixalytics

Following our recent blog about the work we’re doing on machine learning with the MARLISAT project, we were interested to see two different applications using machine learning classification algorithms in the news this week.

France

Reports highlighted a project undertaken by the French tax authorities using a classification solution to identify undeclared swimming pools in parts of the country. Developed by Google and Capgemini, the system used machine learning algorithms to identify the outlines of swimming pools from aerial images, and then cross-referenced these identified pools with those that had been declared to authorities. In France, swimming pools must be declared to the authorities as they attract higher property taxes.

During the trial undertaken in the last quarter of 2021, over nine regions of the country, 20,356 undeclared pools were indentified with the increase believed to have been driven by people working from home over the last few years. It is reported that this exercise has raised an additional €10m in tax revenue, although the tax authorities said they are only encouraging people to register and pay taxes at this stage. It is believed that, nationally, this could raise close to €40m in additional revenue each year. The system is expected to be rolled out nationally, and could be extended to identify other undeclared home extensions, patios or gazebos, which also attract further property taxes.

It’s interesting that it’s also reported that the system had a few issues, with incorrect classifications; elements such with pools obscured by trees not being picked up, or mistaking solar panels, tarpaulins, or industrial farming tanks for pools. We recognise such issues from our own work, when looking for plastic waste we found items made of plastic, such as artificial football pitches, but obviously there were not waste! Accuracy can improve as more training sites are added to the system.

Interestingly, this is not the first attempt to identify undeclared pools. A decade ago, Greek tax authorities used satellite imagery to identify 16,974 private swimming pools in Athens, when only 324, equivalent to 1.9%, had been registered and paid the relevant tax.

Brazil

Also, we noticed a paper published in Conservation Letters, a journal for the Society of Conservation Biology looking at using machine learning and satellite images to combat deforestation in the Amazon. The paper by Mataveli & Oliveria et al., titled ‘Science-based planning can support law enforcement actions to curb deforestation in the Brazilian Amazon’, identified that the current high priority areas being monitored by authorities to combat illegal deforestation could be reduced by 160,000 sq km, enabling authorities to focus on the key areas and make better use of resources. The current strategy, known as the Amazon Plan 2021/22, identified 11 areas it noted accounted for 70% of the total deforestation in the Amazon.

Using applications on TerraBrasilis, an online platform developed by Brazil’s Instituto Nacional de Pesquisas Espaciais, the researchers found that the high priority deforestation hotspots accounted for 66% of the average annual deforestation rate, the 11 areas in the Amazon Plan only represented 37% of this amount with the others being new areas for deforestation

The study used the PRODES tool to look at deforestation monitoring and the DETER tool to look at forest change cover, and applied a Random Forest machine learning algorithm to predict deforestation hotspots and classified areas as high, medium or low priority for deforestation detection.  TerraBrasilis is a freely available tool to enable anyone to review and monitor what is happening.

Conclusion

The potential for machine learning classification techniques using satellite data are becoming more and more obvious, particularly to authorities struggling with costs who can see this as a good way of monitoring what is happening without having to send people physically out to investigate. While these solutions are likely to grow, it is important that the limitations on accuracy are also recognised and the human element is not taken entirely out of the process.