SDG 12: Responsible Consumption and Production
SDG 15: Life on Land
2. Project Details
Company or Institution
dida Datenschmiede GmbH
General description of the AI solution
ASMSpotter is a machine learning and AI software to automatically detect and monitor artisanal and small-scale mining (ASM) on satellite imagery using novel computer vision techniques. ASM is the source of livelihoods for more than 44 million people across 80 countries worldwide, and thus has huge potential to contribute to the achievement of the SGDs. At the same time, ASM can have immense negative impacts on human rights, development and the environment if not governed properly. Effective, continuous monitoring of ASM activity is a crucial component in addressing these challenges. ASMSpotter provides an efficient and effective AI solution by combining cutting edge machine learning with specialist expertise on ASM through the partnership between Dida and Levin Sources.
ASMSpotter is a proven AI solution, piloted on artisanal gold mining in Suriname’s rainforest. A convolutional neural network was trained on a dataset consisting of over 15.000 labelled satellite images from Planet Scope. The dataset was created by experts from the Aachen University. By automating the analysis of satellite imagery, we lower the costs of monitoring and enable producing countries to more effectively govern their mining sector, enabling the protection of large geographical areas such as the Amazon.
ASMSpotter also works with open-access data and satellite imagery, making the tool accessible to Governments or organisations with limited resources. The AI tool achieves comparable performance on lower spatial but higher spectral resolution imagery from the ESA Sentinel-2 satellites. We continue to develop the ASMSpotter, including the expansion of training data towards more regions to increase the accuracy of the machine learning model, as well as data fusion by including Sentinel-1 synthetic aperture radar data to increase the update frequency and thereby make quick reactions of ground teams possible. Furthermore, we are currently working on multiclass labelling to analyse the context of ASM sites.
dida Datenschmiede GmbH
Excellence and Scientific Quality: Please detail the improvements made by the nominee or the nominees’ team or yourself if your applying for the award, and why they have been a success.
The AI architecture of ASMSpotter is U-Net-based, which is a fully-convolutional image segmentation neural network with an encoder-decoder structure. The encoder sequentially applies convolutions to the input image and shrinks the representation of the image, which is called down-sampling. The sequence of convolution and down-sampling is applied repeatedly. The decoder applies convolutions and up-sampling mirroring the effects of the encoder. The up-sampling is facilitated by so-called skip connections. This architecture allows the model to learn structures on multiple scales in the training data at the same time.
The original training data was based on PlanetScope satellite imagery, which is four-channel with a resolution of approx. 3m per pixel. Migrating towards the Sentinel-2 constellation as a source dropped the resolution to 10m per pixel. The higher number of channels (12) compensated for the more coarse resolution and we achieved similar results.
Furthermore, it was helpful to extend the channels by the Normalized Difference Water Index (NDWI) and Normalized Difference Vegetation Index (NDVI). These additional channels highlight water bodies and vegetation in the image respectively.
We obtained an initial F1-score of 0.74. By tweaking the models (adding hypercolumns, pyramid pooling, etc.) and systematically expanding the training data we currently achieve scores above 0.82. (Note that due to the nature of ASM sites having fuzzy and transitory boundaries, the training labels itself inherently prevent a theoretical score of 1). We are also working on new ways to measure ASM site detection success to make a fair comparison between different ML models and application regions possible.
The technology is ready for application, as seen in our Demo. We presented the ASMSpotter at several conferences like the German Aerospace Center’s Copernicus Forum. Currently we are in discussions with several partners from the public, private and civil society sectors to apply the solution.
Scaling of impact to SDGs: Please detail how many citizens/communities and/or researchers/businesses this has had or can have a positive impact on, including particular groups where applicable and to what extent.
ASMSpotter can be used by Government, civil society, and private sector actors to effectively and continuously monitor ASM over large geographic areas. The AI solution is applicable globally, and has proven its accuracy especially in rainforested areas and in detecting alluvial gold mining in and around waterways. Such ASM activity often has large negative impacts on soil, forests, water and air. The solution offers the flexibility to adapt it to particular needs of the user (e.g. overlay with mining cadastres, integrating it into inspection databases, etc) and can be integrated into other existing systems used by the client. Therefore, each application of the tool will have its specific objectives, with the corresponding success indicators, which are jointly developed with the user/client. Each application/implementation project of the tool enables the further development of the tool.
ASMSpotter can thereby support the process of bringing ASM into the formal economy, with the aim of addressing its negative social, human rights and environmental impacts, while harnessing its benefits for livelihoods, poverty reduction and economic development for the local population. Through this, ASM Spotter contributes to the achievement of a variety of SDGs (see above), especially in mineral producing countries, for ASM communities, or for communities that are affected by it, such as for example indigenous communities. In particular, ASMSpotter can help protect human and natural heritage by allowing Governments to detect new ASM activity and compare this with other land uses such as indigenous lands, protected areas or forest reserves, areas with high biodiversity value and protected species. This allows Governments and other users to have a better overview of ASM activity in real time, intervene in a timely manner and manage the sector better.
Scaling of AI solution: Please detail what proof of concept or implementations can you show now in terms of its efficacy and how the solution can be scaled to provide a global impact ad how realistic that scaling is.
Artisanal mining does not only occur in the Amazonian rainforest, where the tool was originally developed. In fact, almost every tropical rainforest is already touched or threatened by some form of small-scale mining and the mining methods used in Suriname are applied also in subtropical regions like central Africa. ASMSpotter, which was developed to identify that form of mining, is therefore easily applicable on a global scale. The architecture of the AI model is also capable of being trained on other forms of mining.
One major challenge we are facing when expanding the covered area is generating ground truth data for new regions. To gather that data, time consuming and expensive on-ground expeditions are necessary. This is where our partnership with Levin Sources comes into play. Levin Sources is specialized in ASM value chains and has a global network of partners, associates and clients in the sector. Jointly we are already in talks with several governments, NGOs and companies to embed ASMSpotter into real world applications.
The machine learning algorithm and the training set were jointly developed with academia, and we continue to include it in new universitarian cooperation to use synergies with other research efforts. One of our projects together with the TU Munich focuses on meta-learning to ease generalization of models, ASMSpotter is one candidate to apply the research.
All in all, ASMSpotter is a beacon project for dida and Levin Sources, fostering the interest in AI based solutions to monitor and protect endangered environments and people's livelihoods. One example of this is winning Microsoft's AI4Earth award last year. The prize increased our visibility and directly led to a project proposal for a Brazilian NGO on monitoring the health of a river system impacted by ASM. Additional prizes would probably have similar impacts.
Ethical aspect: Please detail the way the solution addresses any of the main ethical aspects, including trustworthiness, bias, gender issues, etc.
The partnership between Dida and Levin Sources was established with the explicit goal of including ethical and human rights considerations in the application of the AI tool. This is especially important since ASM communities and the communities impacted by ASM are often the most vulnerable and marginalised already, and include sub-groups such as women, children, indigenous peoples, and others who face intersectional vulnerabilities.
Therefore, the ASMSpotter as a product does not only include the AI technology, but comes as a package that contains advisory services by Levin Sources with the aim to guide the user/client in how the technology should be applied and how the information obtained through it should be used to best foster developmental outcomes in the ASM sector. This includes the development of usage safeguards to be included in contracts with users/clients (to ensure legality and ethical use of the tool by each client). In addition, “governance modules” are offered to users/clients, which aim at supporting the triangulation and verification of data from the tool (i.e. through building ground-truthing systems), the analysis of the data produced through the tool, as well as the development of appropriate policy strategies and responses based on that data.
Through the partnership, ASMSpotter is more than the AI solution developed by Dida, and benefits from Levin Sources’ long-standing expertise and knowledge of good practices in both analysing and governing the ASM sector.