Disclaimer:
Please be aware that the content herein has not been peer reviewed. It consists of personal reflections, insights, and learnings of the contributor(s). It may not be exhaustive, nor does it aim to be authoritative knowledge.
Learnings on your challenge
What are the top key insights you generated about your learning challenge during this Action Learning Plan? (Please list a maximum of 5 key insights)
1. Often illegal dumpsites are non-randomly distributed and concentrated in areas with certain predictive factors.
2. The computer vision (AI) model can be adapted to new locations with little effort.
3. Incorporating Human-in-the-Loop processes can enhance the transfer learning (a ML technique that
enable adapting knowledge from other locations to local context).
4- Enabling open architectures (e.g. digital public goods) can enable the effective adoption of data-driven results by stakeholders (governments, communities)
Considering the outcomes of this learning challenge, which of the following best describe the handover process? (Please select all that apply)
Our work has been picked up by UNDP or the government and has now expanded geographically in our country, Other
Can you provide more detail on your handover process?
The current version of the computer vision (AI) model deployed in undergoing a scaling process to a local authority in charge of protecting the natural and water resources in Lake Atitlan basin, one of the most iconic locations in the country.
Please paste any link(s) to blog(s) or publication(s) that articulate the learnings on your frontier challenge.
Data and Methods
Relating to your types of data, why did you chose these? What gaps in available data were these addressing?
The computer vision model (AI) was trained on a combination of the AerialWaste dataset and the ground truth data from Guatemala. The validation dataset consisted of 208 samples, with 140 negatives and 68 positives.
- AerialWaste dataset that was used to pre-train the classification model is an open-source Italian aerial dataset developed by the Politecnico di Milano (Polimi).
- Ground truth data from Guatemala was used encompassing around 2,000 locations collected in 2021 and 2022 by local authorities.
- Google Maps Static API was used to collect a comprehensive set of aerial images for both positive and negative sample locations within Guatemala.
Why was it necessary to apply the above innovation method on your frontier challenge? How did these help you to unpack the system?
The computer vision (AI) model, based on convolutional neural network (CNN) architecture and transfer learning, provides a cost-efficient method utilizing satellite imagery for detecting illegal dumpsites in Guatemala and hence protecting water resources from solid waste and toxic components.
The performance results of the current AI model show promising trends, especially in terms of the generalization capabilities indicated by the validation dataset metrics.
Partners
Please indicate what partners you have actually worked with for this learning challenge.
Please state the name of the partner:
SDG AI Lab
What sector does your partner belong to?
United Nations
Please provide a brief description of the partnership.
SDG AI Lab partnered with UNDP Guatemala Accelerator Lab for identifying and testing cutting-edge tools and techniques like machine learning and artificial intelligence combined with advanced GIS and remote sensing techniques.
https://sdgailab.org/
Is this a new and unusual partner for UNDP?
No
Please indicate what partners you have actually worked with for this learning challenge.
Please state the name of the partner:
Central and local government
What sector does your partner belong to?
Government (&related)
Please provide a brief description of the partnership.
The Ministry of Environment and Natural Resources provide insights and statistics to inform the co-creation and technology selection process.
Local government (such as AMSCLAE - Authority for Management and Protection of Natural Resources of the Lake Atitlan basin) provide feedback and a test environment for the current version of the computer vision (AI) model.
Is this a new and unusual partner for UNDP?
No
End
Bonus question: How did the interplay of innovation methods, new forms of data and unusual partners enable you to learn & generate insights, that otherwise you would have not been able to achieve?
Despite challenges related to lack of high quality statistics, traditional methods (e.g. surveys) for collecting on the ground information, and low knowledge on the potential benefits of emerging technologies, the interplay of co-designing solutions with local authorities and global experts leveraging state-of-the-art analytical methods and low-cost imagery led to the deployment of a promising computer vision (AI) model for the benefit of local authorities and communities. These results have been recognized in global events such as the recent UN World Data Forum held in Colombia during November 2024 (see session TED.201 at https://unstats.un.org/unsd/undataforum/programme/)
Please upload any further supporting evidence / documents / data you have produced on your frontier challenge that showcase your learnings.
Result presentation showcased at the UN World Data Forum, Colombia, November 2024
The closing form saves automatically or via the blue "save changes" button the top left. Thank you
11Sustainable cities and communities
13Climate action
17Partnerships for the goals
Comments
Log in to add a comment or reply.