Waste management is a critical issue from an environmental, social and economic point of view. Global studies predict that by 2025, the cost of solid waste management will reach $375.5 billion annually [1]. Since the 1980s there has been a significant increase in waste disposal costs and consequently a greater incidence of illegal practices managed by organized crime. According to a 2015 report [2], the number of illegal landfills in Europe is 12,628 for a total of 2,871,186 tons of waste. The problem affects all EU countries. Italy, Poland and Germany are the countries with the largest amount of illegal waste. Europol reports illegal waste trafficking as the biggest growth area for organized crime, with revenues in the billions of euros annually [3].
The PERIVALLON project and geospatial intelligence
The PERIVALLON project is partially funded by the European Union’s Horizon 2020 program under grant agreement No. 101073952. The project includes 5 Police and Border Guard Authorities, 3 environmental protection agencies, 6 Research/Academic institutions, 8 industry partners, one EU Agency, and one Foundation, and the achieved results will be will be validated in field tests in four operational use cases in different European countries.
PERIVALLON aims to introduce Geospatial Intelligence and Earth Observation from satellites, airplanes and drones to counter illegal waste management through more effective identification of potential breaches to environmental regulations. Geospatial Intelligence helps detect violations in authorized waste treatment plants and also totally illegal landfills. PERIVALLON exploits a two-level strategy. First, a periodic large scale mapping of the territory is carried out by means of Earth Observation from satellite and aircraft, Geospatial Intelligence and Artificial Intelligence. This phase provides a list of potentially critical sites. These are evaluated by the relevant institutions to select those to be subjected to inspection, during which drones help mapping the types of waste and an assess the quantities and risk level.
Artificial intelligence at the service of geospatial intelligence for the identification of illegal landfills
Automatic image analysis is a sector of information technology that applies computer vision algorithms to solve many problems: from industrial robotics to autonomous driving to environmental surveillance. In recent times, the traditional methods of image analysis, which required ad hoc procedures for the extraction of useful characteristics on which to base the analysis, have been replaced by approaches based on machine learning (Machine Learning). In particular, Deep Learning based techniques have demonstrated impressive performance in many data analysis problems and especially in image analysis. A turning point is represented by the advent of convolutional neural networks (CNN). A CNN is a data processing architecture that relies on the convolution operation to condense input information and produce a compact representation that contains only the information needed to compute the desired output. CNNs are used primarily for image classification and object detection. Image classification predicts the types of objects that appear in an image whereas object detection determines the region of the image in which a specific type of object appears. PERIVALLON researchers apply CNNs to satellite image analysis to automate the identification of illegal waste deposits. The choice of CNN fell on the ResNet50 architecture [4], a 50-level CNN that provides an excellent compromise between classification accuracy and computational power required for training. A knowledge transfer procedure (Transfer Learning) was applied to ResNet50, which consists in subjecting a network trained in the classification of generic images to a partial retraining with aerial images. In this way the training process is accelerated and a few hundred satellite images classified by ARPA Lombardia experts are enough to achieve high precision.
The designed CNN can identify sites of potential interest in satellite images with an accuracy of 92%. The following figure shows two examples of images classified from the realized CNN.
Examples of automatic identification of potentially critical sites using the designed CNN architecture. On the left the original images. On the right, the “Class Activation Maps”: they represent the areas on which the algorithm focuses for classifying the site as critical.
Thanks to PERIVALLON, it is now possible to automatically analyze a vast territory in a few hours, a task which requires weeks if done manually. The automatically produced lists of locations are sorted according to the probability that a suspicious site is present, so that the operators in charge can focus on the most “promising” sites first.
[1] Dan Hoornweg and Perinaz Bhada-Tata. “What a waste: a global review of solid waste management”. In: Urban development series knowledge papers 15 (Jan. 2012), pp. 87–88. url: https://openknowledge. worldbank.org/handle/10986/17388.
[2] Emma Watkins. “A case study on illegal localised pollution incidents in the EU.
A study compiled as part of the EFFACE project”. London: IEEP (2015) https://efface.eu/sites/default/files/EFFACE_Illegal%20localised%20pollution%20incidents%20in%20the%20EU.pdf
[3] Europol. Trash worth millions of euros. 2019.
https://www. europol.europa.eu/newsroom/news/trash-worth-millions-of- euros
[4] HE, Kaiming, et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-778.