​

Overview

​

AI Climate is a decision making tool for climate-resilience planning geared to secondary and tertiary cities and their regions which lack data at the local level and have limited technical capacity and resources. It uses machine learning (ML) technology that processes geospatial imagery and georeferenced datasets, deriving data in the form of analytics-ready layers such as impacts of flooding, landslides, land degradation, and others specific to the region. With these layers, AI Climate maps, monitors, assesses, and predicts climate change-related impacts in the city, while focusing on the poorer, most vulnerable communities within them. The AI Climate platform is ready to use, faster, and cheaper to update compared to traditional methods. It is free and available to public officials and communities for decision making. Most importantly, through a training and crowdsourcing component, AI Climate tool empowers local and informal communities to co-design equitable, nature-based, and ecosystem restoration solutions. The AI Climate platform is underpinned by an open-source strategy, and a system scalable in size and scope. It is straightforward to replicate in cities within the same region and, with few modifications, to other regions of the world. To the best of our knowledge, the AI Climate tool is a unique mix and match approach of processes and techniques for producing data for secondary and tertiary cities. (2020)

​

Context and Hypothesis Origin

​

In 2018, the World Bank estimated that almost 30% of the world’s population lived in informal settlements and other vulnerable urban communities. These communities are expected to expand rapidly in the Global South (World Bank 2019) over the next decades, especially in secondary and tertiary cities and regions, which are the gateways for rural-to-urban migration and displaced populations. As a result, these urban areas, where about 75% of the world population lives (Cities Alliance, 2018) are undergoing the fastest processes of urbanization and informal settlement formation. They do not, however, receive the attention and funding devoted to primary cities, such as national capitals. Concurrently, the escalation of global warming and resulting climate change impact is generating new and more extreme risks (Intergovernmental Panel on Climate Change, 2021) for these expanding communities. These forces combine to create an urgent need to plan for more proactive and efficient resiliency strategies to prevent climate hazards from becoming disasters. Yet secondary/tertiary cities and their regions are impeded from doing so because they lack the necessary local data granularity and technical resources. Excluded communities within these cities will thus suffer disproportionately from the worsening consequences of climate change in the absence of any action.

​

In recent years, we saw an opportunity to address the lack of local data, high costs for resilience-planning and slow pace of adaptation to rising climate risk in secondary and tertiary cities and regions.

​

Our hypothesis is that with increasingly abundant spatial data, and rapidly falling processing costs, advances in Machine Learning (ML) creates the opportunity to drastically cut the time and expense involved in hazard mapping, assessment and prediction, freeing funds to build resilience in cities and communities. Such a combination of data and ML can facilitate visual accessibility to areas otherwise inaccessible due to budgetary, geographic, or security constraints; transform current urban planning processes by reducing their time and cost, while increasing their transparency, leading to effective risk reduction; and allow updates of local data and thus of risk assessments at a frequency that does not exist at present, crucial for hazard prevention and planning at a time of rapid urban growth and rapid change in climate patterns.

​

Proof of Concept (PoC) Beta Version 1.0 (link here)

​

To explore this hypothesis, since 2019 we have been developing AI Climate. A concept paper was accepted for presentation at the 2020 World Bank Land and Poverty Conference. We followed with a proof of concept (PoC) focused on river/rain flooding in two cities in Honduras, which we completed about six months ago together with our technology and scientist collaborators in Honduras. The PoC showed that ML algorithms can identify river/rain flooding risks and socially vulnerable communities (more detail in data maturity section). In April 2021, we were awarded a Microsoft AI for Earth grant, which provides us with significant free data storage and processing capacity for the next year. When COVID protocols permit, we will add a crowdsourcing component that incorporates knowledge from the local community, and a participatory scenario planning component to develop strategic, nature-based, ecosystem restoration interventions to mitigate local current and expected future climate hazards through the participatory development of local interventions.

​

During our PoC in Honduras, we used CNN, U-net architecture, a proven standard of the industry. It involved the development of ML algorithms to identify socially vulnerable urban areas, to predict their spatial growth patterns (recurring CNN) and to identify areas at risk from flooding (hybrid models for flooding susceptibility). Input layers for ML training for rain/river flooding include aspect, slope, elevation, terrain water absorption, and flow accumulation. The combination of these different layers and algorithms enabled us to identify current flood hazards throughout the city, focused on those faced by socially vulnerable urban areas, and to predict future hazardous land as well as how settlements will evolve. The final output is an up-to-date flood-risk heat-map applicable to the city and its region, featuring the footprints of socially vulnerable areas at present and in the proximate future.

​

We are proud to state that, while developing our PoC for river and flash-flooding layer feasibility in Honduras in 2020, we were able to provide “just in time” information (in the form of a predictive flooding risk map) to our Honduran partner IHCIT for the city of Tegucigalpa during last year’s catastrophic flooding season in the country.

​

Our ML component combines optical and microwave remote sensing techniques. The PoC was conceived to demonstrate that ML can obtain plausible and valuable results (averaging 80+% certainty) using data in the public domain. We tested, compared and retrained both free high- def-model (satellite images (res 0.40 meters, 2013)) with lower-def-images (Sentinel 2) for better results; although lower-resolution image delivered less local details. We are now in the process of improving our algorithms.

 

​

​

Alignment with our Mission

​

For the past two decades, as urban and regional planners at the Institute for International Urban Development (I2UD), spun off from the Harvard Graduate School of Design in 2005, we have pioneered strategic planning centered on climate change adaptation, social inclusion and informal settlements, and local capacity building throughout the Global South. Our mission is “to formulate sustainable urban and regional development strategies with local governments and communities building upon their unique cultural and social assets.” To preserve these local assets, it is our mandate to co-design with local partners; thus local crowdsourcing intelligence and local data validation is of paramount importance.

​

This project furthers our mission in consonance with our expertise and track record, as it leverages cutting-edge data and data processing technologies to help cities and communities plan for a sustainable future in the face of climate change. We seek data production to provide concrete, more climate-resilient, inclusive and equitable planning outcomes, particularly to foster climate change adaptation and mitigation through an emphasis on nature based, green infrastructure, and ecosystem restoration solutions with benefits for an entire city and region, and focused on the most socially vulnerable populations.

​

​