Analysing and visualising the efficacy of international climate initiatives (COP26) through data
Our Rationale: In a fast-paced world governed by the forces of data, as much as many of our problems are manifest from the pressures that technology has put on us; it’s clear that technology also serves to inspirit the very solution of some of the problems that our rapidly developing world has served to create.
And, of course, there is no example where this is quite as true as climate change.
We all know that climate change has serious implications for the health & futures of young people; but I think I speak for all of us when I say that when we young people look at ways we can actively combat it, we find ourselves quite powerless. As such, we’ve seen a rise in a phenomenon that is Climate Anxiety.
As the chart shows on the left, we have seen a rise in the number of young people across different geographies expressing a fear of the future of the environment; and as two ecowarriors ourselves, we found ourselves particularly compelled to examine how our interests in preserving the environment intersected with our interests in Data Science that we had developed over this term. Upon further research, it became clear that the science, policy and communication practices around DS will impact the solutions we utilise into the future with climate change. Data science is a phenomenal tool in helping climate researchers really understand the uncertainties and ambiguities inherent in climate data to ultimately identify interventions, strategies and solutions to fulfill the needs of both humanity and the environment - assisting loads in cases where decision makers are often faced with numerous and conflicting goals. These strategies empowered by data and data science were exactly the sort of thing we hear of at the likes of COP27, where our very own Minouche Shafik spoke in November 2022. And is where pledges determined by data like those shared at COP27, involving the likes of countries promising to limit global warming to 1.5 degrees are predicated. Data science is what is used to examine the uncertainty of climate models and ultimately help visualise the issue at hand.


It is in this vein that we have aimed to use Data Science to go one step beyond this and, in an almost self reflective way, examine the efficacy of the very initiatives and pledges that Data Science has helped to create and will do so by using an Air Pollution API. We will also go one step beyond this and examine how the degree of response to climate change following a global climate protocol like COP26 varies depending on the level of economic development of countries as well. Preliminary Targets: Our initial targets following our complete change of project centred around investigating the efficacy of the climate initiatives. We have seen, over the last few decades, countless environmental initiatives being introduced. We were very interested in learning exactly what effect these initiatives had on the pollution levels of different countries. Given our API restrictions (will be examined further), we decided to focus on COP26. We are interested in answering the following questions to help us further examine the efficacy of climate initiatives (COP26) and, moreover, ask ourselves exactly what this means for the environment:

• Why have so many different climate initiatives been introduced over time?
• Did COP26 in particular turn out to be successful?
• What did the initiative have the largest impact on?
• Do different countries respond differently to these initiatives?
• Do discrepancies in the response of emerging and developed nations exist?
• What impacts have these initiatives had on the Air Quality Index and other polluting gasses?
• What time lags are associated with these initiatives?
• What are the short, medium and long-term effects of COP26 on the environment?

Cluster containing Newly Emerging Cities: Delhi and Gaborone

Cluster containing Low Income Cities: São Paulo and Delhi

It is important to note that we initially wanted to explore the impact on Beijing, but inputting the longitude/latitude into the API kept generating an error, and we decided instead to investigate the effects of COP26 on the pollution in São Paulo instead - which is in a similar

The choice of our 6 cities was entirely orientated around the fact that we wanted to have a dimension of our findings that examined the extent of environmental response, and how this depended on the level of economic growth that was encountered by a country. Our 6 locations allowed us to gather good intel on a range of countries that were epitomes of their economic categories.

Generally, economic theory produces 5 fundamental reasons why the level of environmental response might vary by how wealthy a country is:

1. Financial resources: Wealthy countries typically have more financial resources to invest in environmental protection and sustainability initiatives. They can afford to implement more expensive and advanced technologies to reduce pollution and protect natural resources.

2. Economic priorities: Developing countries may prioritise economic growth over environmental protection, as they may see environmental regulations as a hindrance to economic development. On the other hand, developed countries may have a stronger emphasis on environmental protection as their economy has already been established.

3. Political will: Wealthy countries may have a stronger political will to address environmental issues, as they have the resources and stability to implement policies and regulations to protect the environment. Developing countries may have less political will to address environmental issues due to lack of resources and other pressing concerns.

4. Public awareness: Developed countries generally have higher levels of education and more access to information, which can lead to greater public awareness and concern about environmental issues. This can lead to greater pressure on governments and businesses to take action to protect the environment.

5. Technological capabilities: Wealthy countries typically have greater technological capabilities to address environmental issues, including the ability to monitor and assess the state of the environment, the ability to develop and implement new technologies to reduce pollution and protect natural resources, and the ability to adapt to the impacts of climate change.

We want to see how these fundamental tenets in the relationship between level of economic wealth and responsiveness to climate change can be visualised and scrutinised with data science. Our choice of these 6 cities enabled us to do exactly this. Furthermore (and perhaps more fundamentally) when it came to deciding which countries to use as part of our analysis we first identified the signatories of COP26. All of these countries had, to different extents, committed to the conference and made pledges. Our Choice of Further Variables (Dates and Pollutants): Dates (or, perhaps more accurately, our lack of them) were wholly based on our limited choice of data sets - we will examine this further in the next section. As for our choice of pollutants, we had quickly discerned that we required a range of different pollutants in order to avoid the impacts of any selection bias that could have arisen from selecting only one pollution metric. Descriptions for our chosen pollutants are as follows:

• The first polluting gas we looked at was Carbon Monoxide. Carbon monoxide occurs primarily from emissions by fossil-fuel powered engines. We decided to look into Carbon Monoxide as one of our pollutants of interest as the API was able to return data on it. It seemed natural to focus on the evolution of this gas with the advent of the climate initiative
• We then decided to look at Ozone/ Oxygen Trioxide. When inhaled, ozone causes severe damage to the lungs. Ozone absorbs radiation, consequently acting as a greenhouse gas. As such a big contributor to the greenhouse effect, we felt it logical to investigate its evolution,
• NO2 or nitrous dioxide occurs primarily from cars, buses and trucks. It gets into the atmosphere, reacting to form Nitric acid (acid rain). This acid rain causes damage to buildings and causes the acidification of water bodies, contributing to biodiversity loss.
• We also examined the fine particulates (PM2.5) - these appear in the air, reducing the visibility. Exposure to PM2.5 has been said to cause premature mortality - so we chose to investigate this as our fourth metric of interest. Potential Datasets: We found out a lot of data was available online. Below are the sources we found, with the respective numbers of datasets that each source held:

1. Open Weather Map’s API
2. WHO GHO OData API
3. European Environment Agency had 227 datasets available for use
4. Data.gov had 12 datasets available for use
5. Data.gov.uk had 7 datasets available for use
6. UKCOP26 had 4 datasets available for use
7. Local Government Associate had 5 datasets available for use

Making Use of Code: To make our code usable, we initially had to import in all of the libraries (requests, pandas, numpy, json and matplotlib.), these gave us the critical tools we needed to ensure that we were able to access all the tools needed in order to execute all required functions to our data. We started by analysing OpenWeather’s Air Pollution API, seeing the total 9 variables and began to decipher which ones could be used to universally answer our question. For some obscure air pollutants like PM10, we had empty values for most countries so we had to withdraw the prospect of any use of these. Eventually we settled on using the pollutants Nitrous Oxide, Carbon Monoxide, Ozone as well as PM2.5 particulates as our variables of interest. We created this flowchart that can be seen below, to illustrate the scale, size and hierarchies involved with really finding ourselves with the information that we wanted to illustrate in our visual analysis: The OpenWeather API has an almost infinite amount of air pollution data that provides information on a daily frequency over the past 2 years from all across the globe. We found that each country had approximately 17,500 rows of data which across our 4 chosen variables postcleaning and 6 countries meant that we were processing calls and requests and dealing with approximately 945,000 individual data points.

When we break this down into our clusters of 2 countries each, separated by their level of economic development; low income countries, newly emerging economies and high income countries found themselves with paginated data that was approximately equal in size at around 315,000 data points.

And then when splitting our 2 year time frame into 3 distinct and equal length periods of beforeCOP26, during-COP26 and after-COP26 we found ourselves with data points across our 4 chosen pollution metrics at about 105,000. Which surprisingly, at least compared to other groups, was a lot more manageable than it had initially seemed.

We utilised the drop function to allow us to drop unnecessary columns. This allowed us to drop the columns of the variables that we were not actually interested in (SO2, CH4). Using the concat function, we merged the 6 dataframes into one, main, dataframe. This made subsequent data operations a bit easier. We also utilised the .loc function to filter out the time range that we needed (pre, during and post COP26). Use of Pandas: We used the pd.Dataframe function to create the data frame itself. We then made use of several pandas functions such as concat (to merge the data frames together) and drop to remove unnecessary columns. We utilised df.rank to do dataframe operations - finding which countries were the largest contributors to pollution before the inception of COP26 and how these rankings evolved with the advent of the initiative. Missing data: We had to employ the use of the DropNA function for this. This allowed us to remove all NaN values that, come data visualisation, would not be able to return anything meaningful and would have produced many annoying errors. You’ll see this in practice in some of the correlation data we will employ for Gaborone’s during and post COP26 data.

Joining Several Databases: We used .concat which concatenated the 6 different data frames into a singular dataframe. We then placed these multiple data frames into a singular list. The diagram towards the bottom on the left illustrates this process. Preliminary Visualisation (Time Series): Briefly explaining some of the code we employed in the experimentation we started off with; MatPlotLib provided the basis of our preliminary experimentation. This is something that we have already dealt with through problem sets and on the few python courses we have done ourselves. But, in essence we started off by plotting graphs of the same pollutant (on this page you can see PM2.5) for our 6 different cities and timeframes; that being before, during and after COP26. Obviously plotting the designated start and end times for these periods on the x-axis and the pollutant levels on the y-axis, we were able to access quite a comprehensive graph illustrating air pollution levels over the course of a period for all cities and potentially a bit of a story associated with this data. The first graph demonstrates the PM2.5 levels before the advent of COP26. We can see huge fluctuations in the level of PM2.5, especially with Delhi. Looking at the graph, it is very hard to discern between the PM2.5 emissions of the HIC; they are covered by the other clusters, so we decided to use a slightly different structure when we mapped out the PM2.5 emissions during COP26 (second graph). We still see big fluctuations in Delhi’s levels of emissions, however there is a visible and significant reduction in the emissions of São Paulo. This can be explained by the firm commitment that Brazil made during the conference. After the advent of COP26, Brazil’s emissions have reduced significantly compared to those before the introduction of the initiative. Delhi seems to continue to have significant PM2.5 emissions - due to the soft commitments made in the short term. Ozone emissions, similar to PM2.5, seemed to have large variations across the board in the build up to COP26. Once again, India seems to be a big contributor, with Sao paulo and gaborone also remaining big emitters of oxygen trioxide. COP26 seems to have a far more significant reduction on Ozone emissions compared to PM2.5

• across the board, ozone emissions have dropped significantly for the different nations, with significant drops for Port-Au-Prince and Gaborone - the conference ended up having significant decreases in 03, and a comparatively larger impact than PM2.5 After COP26, the emissions have decreased across the board, but have increased slightly compared to during COP26. This reflects almost a ‘wearing off’ of the effects of the conference, with countries resorting back to their old ways as time goes on. To highlight this aspect of firms resorting back to their old ways, we decided to graph the NO2 emissions during and after the conference - here too, it seems that as length of time since cop26 increases, the effects it had start to wear off slightly - perhaps this is why so many initiatives have been targeted over the last couple decades.

3. After COP26, Delhi found itself emitting similar carbon monoxide levels as pre-COP26. Sao Paulo increased emissions by 83%; the effects seem to wear off with time. More Refined Visualisation (Pie Charts) [O3]:
4. Before the conference there seemed to be an asymmetry in the ozone pollutant levels, with most cities falling around the 20% region.

Our first step of cleaning the data involved us firstly importing in all the libraries (requests, pandas, numpy, json, matplotlib (to plot our data). We used Google Colab as we had to import files (since it operates on the cloud). We then accessed all of this data through the API key we received upon signing up to OpenWeather’s air pollution API. We then collected our data using location - this involved using specific longitude and latitude information to access the data we needed. The lon/lat functions specify the coordinates for each of the cities that we were interested in. We added the start and end dates, to highlight pollution in the build-up, during and post the introduction of the COP26 initiative. Using a Unix time converter, we converted the start and end dates into the unix form. The API Documentation needed the unix code to process the start dates. We then repeated this process for each of the other cities that we were interested in. Step 2 - Cleaning the Data: Using our example of São Paulo, let’s walk you through each step behind the cleaning of the data and then, moreover, how we employed this cleaning to pivot, melt and mutate our data:

1. We used the combine function to make the data frames into a single list and assigned it to the variable ‘combined,’ this allowed us to loop through the data frames; allowing us to execute the column dropping and concatenation as efficiently as possible.
2. For each value in the list of combined, we run a loop to change all unix code to datetime. Drop columns for main, components and dt.
3. Concatenating data frames, and merging them around axis 1 (the second axis) - admittedly, we needed to use our week 8 here.
4. We concatenate for each of the cities, merging the data frames together respectively (step 3-step 8)
5. Step 9 - made a new list containing the new data frames. Defined a new dataframe under the combined2 variable that brings together all the cleaned data frames into a single list.
6. Step 10 - dropping the columns for main, components and dt, this time for our new combined2 variable

As the outcome of this cleaning, which required a process of pivoting, melting and mutating our data, can see when we visualise our same example of Sao Paulo, we have a much cleaner dataframe. The dates for our 17,500 rows for this location over the total time we are able to access are very clear, as are the pollutants we have a particular interest in which find their associated values going down on the left hand side. Initial Heatmap Exploration: What does this clean data mean in terms of hard core data visualisation? In week 8 of the course, we did a lot of work with MatPlotLib in order to produce some basic continuous graphs of air pollutants across our 3 distinct time frames of pre, during and post COP26 for all of our cities. But one of the biggest criticisms we had for ourselves was that we wanted to be a little more creative, we wanted to be a little more advanced.

So we entered the wonderful world of heatmaps. Which, through a process of continuous refinement as we will present over the next few slides, really ended up being a phenomenal way to visualise, compare and contrast air pollution metrics across time periods.

The above here shows some of our initial efforts. We can see that our heatmaps are initially quite obscure, and whilst showed an accurate reflection of variation in all of our pollution metrics across all of our cities across the whole time period provided by the API, we were still unable to get clear information out of this and moreover, information that could then provide the grounds of summary stats in order to really help us answer the question at hand. Further Heatmap Exploration: These more refined heatmaps try to elucidate what the real target areas had been during the cop26 initiative. The X axis includes the days of the week, whilst the Y axis involves the different months. We see that, for all of the countries (excluding Gaborone - the missing data which, as explored through NAN values) didn't return us with a heatmap; we will explore this in Statistical inference. There seems to be the highest pollution in the middle of the year, which ceases slightly towards the latter points of the year. It seems intuitive, that there are short term time lags at least whilst the plans are being introduced. Refining Our Heatmaps: After our attempts, we really bolstered down our efforts to create a much clearer and more representative heatmap. After much persistence, we were able to do this; the following providing the steps: In step number 1, we import the datetime module. This is the heart and essence of our ability to break pollution data by time and date.

In step number 2, we rename our columns so that they make sense.

In step number 3, we proceed to concatenate all 6 of our locational data frames/

In step number 4 proceed to drop NaN values and you can see some of this in step number 5, where when we visualise this dataframe, you can see a slew of 0s for the data concerning gaborone, where we had a considerable problem with missing data.

In our final step, we then use calplot, which is a function that creates heatmaps from pandas time series data. And enables us to structure our heatmaps such that they are colour coded by calendar days. The Outcome Of Our Refinement: Below presents our final and refined heatmaps. It is clear that, through these, we are able to tell a story. This is particularly relevant as we begin to examine the specific measures that were undertaken by each respective country:

1. Washington: Washington, D.C. had moderate levels of air pollution before COP26. During the conference, the US government announced new plans to reduce emissions from the transportation sector. As a result, heat maps show a small decrease in pollution levels in Washington, D.C. during and after COP26.
2. Sao Paulo: Sao Paulo had high levels of air pollution before COP26. However, during the conference, the Brazilian government announced new plans to reduce emissions from the transportation and industrial sectors. As a result, heat maps show a moderate decrease in pollution levels in Sao Paulo during and after COP26.
3. Gaborone: Gaborone had relatively low levels of air pollution before COP26. However, during the conference, the Botswana government did not announce any major plans to reduce emissions. Heat maps show little to no change in pollution levels in Gaborone during and after COP26.
4. London: London had relatively low levels of air pollution before COP26. During the conference, the UK government announced new targets for reducing emissions from the power sector. As a result, heat maps show a small decrease in pollution levels in London during and after COP26.
5. Delhi: Delhi had high levels of air pollution before COP26. However, during the conference, the Indian government did not implement as many strict regulations. Heat maps show only a slight decrease in pollution levels in Delhi during and after COP26. 6. Port-au-Prince: Port-au-Prince had moderate levels of air pollution before COP26. However, during the conference, the Haitian government did not announce any major plans to reduce emissions. Heat maps show little to no change in pollution levels in Portau-Prince during and after COP26.