Fabian.2.Nagel@uni-konstanz.de
Giuliano-Andrea.Castiglia@uni-konstanz.de PRIMARY
Student Team: YES
Tableau
R
Python, Pandas
Knime
Typescript: node.js, d3, turf.js, lodash, leaflet
Postgres + PostGis
Approximately how many hours were spent working on this submission in total?
800h
May we post your submission in the Visual Analytics Benchmark Repository after VAST Challenge 2019 is complete? YES
Video
IEEE VAST Challenge 2019 mc-2 Uni-Konstanz
Application link: cpmViz
1 – Visualize radiation measurements over time from both static and mobile sensors to identify areas where radiation over background is detected. Characterize changes over time. Limit your response to 6 images and 500 words.
In order to distinguish between background radiation and contaminated locations we performed first an outlier detection on all the measurements. The bimodal distribution of outliers for static and mobile sensors led us to the conclusion that contaminated locations in our definition are outliers belonging to the second mode of the distribution. This still holds true even for each sensor type on its own. We can spot a major difference between static and mobile sensors in the image below as the highest contamination measured by static sensor is at 1068 cpm and for mobile sensors it is 2418 cpm. The vast majority of data is measured by mobile sensors as can be seen in the distributions.
InIn our VA tool points shown in the map indicate such contaminated locations measure by mobile sensors. The temporal distribution of the contaminated locations is indicated with ticks in the same color range in the lower timeline view.
TThe timeline view displays a fixed timerange of 24h, we can analyze the development over time and spot locations in the map where areas with high contamination where measured, each row represents one mobile sensor and they are ordered by the amount of data they measured over the whole period of the disaster.
For the second, third and fourth day we can identify that the districts Southton, Weston, West Parton, and Downtown have an increasing amount of contaminated locations.
We introduce a different way to find areas of interest by dividing St. Himark into voronoi cells based on the position of Hospitals, Static Sensors and the Powerplant itself. With this new division we can see that the regions belonging to Trauma Hospital, Eaglepeak Hospital, and Static Sensor 13 are covering the area with a higher density of contaminated locations over the full period.
Analysing the timeline for the whole period we see that earthquake at day 3 in the morning causing nearly all mobile sensors to stop measuring for two hours. On the same day at around 5 pm nearly half of the mobile sensor stopped recording or recorded only very short streak with no contaminated locations at all. This pattern is repeated from Thursday 11pm to Friday 8 pm.
Only sensors 2, 10, and 29 have more stable measurements in that period.
Another pattern that can be detected are the unstable measurements from the sensors recording at the first two days.
2 – Use visual analytics to represent and analyze uncertainty in the measurement of radiation across the city.
a. Compare uncertainty of the static sensors to the mobile sensors. What anomalies can you see? Are there sensors that are too uncertain to trust?
b. Which regions of the city have greater uncertainty of radiation measurement? Use visual analytics to explain your rationale.
c. What effects do you see in the sensor readings after the earthquake and other major events? What effect do these events have on uncertainty?
Limit your responses to 12 images and 1000 words.
Mobile and static sensors generally seem to show different sensor response behavior, resulting in deltas up to 1200 cpm. Below visualizations show measured values by both sensor groups, in which the gaps that were not recorded by static sensors are highlighted.
At this point, it is important to keep in mind that mobile sensors cover a lot more are and thus have a much higher chance of actually measuring contamination that is distributed over the city.
The
static sensor contamination over time are plotted in scatterplot were
we can see changes for each static sensor contamination throughout
the five days.
Our visualization tool uses a timeline view to showcase mobile sensor activity periods, in which the temporal dimension is mapped to the horizontal axis. Over the whole width of the browser, a 1-day time interval is visible and continuously selectable by the user through scrolling. The visible time window acts as a reference for other components, making the timeline our primary component for temporal drill-downs.
A sensor activity period is described by streaks of continuous measurements. In these periods, the sensor manages to provide data every 5 seconds, representing a stable, trustworthy data series.
In the case of a reception loss, the time span to the previous sample exceeds 5 seconds, which causes the current streak to end.
Another streak is formed that continues as long as the optimal 5-second sampling interval is satisfied.
Our approach visualizes sensor uncertainty by displaying these streaks for each mobile sensor in an organized, temporally-aligned manner. Clusters of short measurement periods create visual noise that is easily detectable by the viewer. An individual sensor’s uncertainty can be expressed by its average streak length: The shorter, the more unstable the sensor’s behavior and thus, more uncertain.
Furthermore, our vertically-stacked layout and temporal alignment makes it easy to detect patterns and changes in sensor activity periods and their individual uncertainty. The background color of sensor streak rectangles is mapped to the individual streak’s standard deviation with respect to the sensor’s overall mean measurement. This aspect also shows uncertainty in a way that
highlights altering sensor behavior, especially in conjunction with the already described concept of visual noise.
Concerning the reliability of individual mobile sensors, it is generally a hard task to accurately quantify this for each sensor. Based on visual comparisons of above mentioned techniques, we identify mobile sensors 6, 1, 12, 10, 7, 8, 9, 6, 42, 17, 45, 28, 22 to be on the lower end of trustworthiness.
We also identified static sensor 15 as an anomaly, since it stops gathering data for almost half a day which is unusual for a static sensor that is not affected by cellular network malfunctions. Furthermore, every other sensor continuously records for the whole duration.
As for uncertainty of measurements among different city regions, our tool is equipped with an interactive map of St. Himark. In general, our conception of uncertainty is based on the concept of the “signal to noise” ratio. The more data is available, the better anomalies like sensor reception losses, biased/shifted measurement values or missing data will be smoothed out.
In conjunction with the timeline, the number of measurements per district are counted for a certain selected 1-day timeframe. A darker grey value refers to a larger number of measurements, thus reflecting a rather low amount of uncertainty for that region. On the opposite, brighter gray values showcase areas with rather low amounts of measurements.
In the first two days before the earthquake, we can generally see an even distribution of measurements over the whole city. The south-eastern region generally is a bit more uncertain, whereas there is lots of data around the Old Town district. We can see a general decline in certainty after the earthquake, again especially affecting the south-eastern region and the power plant’s district.
On the last day, we see uncertainty expanding from the south-eastern region as well as from the western part of the city where the proximity to busy districts initially caused for lots of measurements to be available.
The put the districts’ uncertainty visualization into perspective, the detail view shows a bar chart of average streak lengths per district. This metric highlights spatially-conditioned effects on individual sensor’s uncertainty. Districts that show a rather low average streak length contain mostly sensors that are on the lower end of the certainty scale, since they suffer from more frequent reception losses.
To highlight the connection between map and detail view, clicking on any district will highlight the corresponding bar in the average streak length plot. In the picture below it is clearly visible that there can be interesting interplay between the two measures: The first red bar shows the highest average streak length of all districts, while the district itself is on the low-end of our uncertainty scale. This essentially means: The district suffers from sparse data, but the quality and stability of the existing data is very good.
As for events and their effect on sensor data, the most obvious one is the earthquake that can be clearly seen in the timeline on Wednesday (08/04) from around 8:30 am to 10:40 am. Almost all mobile sensors start to struggle from reception loss, thus resulting in a greatly increased overall uncertainty.
There is also a drastic change in measurement counts per district after the earthquake, most likely resulting from malfunctioning mobile sensors due to power outages and reception losses. This is also reflected in the average streak length, which drops from roughly 24759 to 15577 after the earthquake.
3 – Given the uncertainty you observed in question 2, are the radiation measurements reliable enough to locate areas of concern?
a. Highlight potential locations of contamination, including the locations of contaminated cars. Should St. Himark officials be worried about contaminated cars moving around the city?
b. Estimate how many cars may have been contaminated when coolant leaked from the Always Safe plant. Use visual analysis of radiation measurements to determine if any have left the area.
c. Indicated where you would deploy more sensors to improve radiation monitoring in the city. Would you recommend more static sensors or more mobile sensors or both? Use your visualization of radiation measurement uncertainty to justify your recommendation.
Limit your responses to 10 images and 1000 words
Considering the uncertainty observed in question 2, there are definitely measurements that are reliable enough to locate areas of concern. The most alarming of all measurements happened on Thursday between 6 and 7 pm by mobile sensor 29, which did not change location at this point. The cluster is clearly visible on the map and the measured values reach from the lower end of the spectrum to one of the highest existing measurement values. While it is embedded in the very uncertain south-east region, it is embedded in a reasonably-sized measurement streak. Plots of average streak lengths for this region also show that the few measurements that exist tend to be more reliable than expected at first sight (see Task 2).
Because of the sudden appearance of this cluster, it might be possible that a moving contamination like a car was placed next to it, while the sensor itself remained static. The magnitude of the measurements in comparison to others suggest that this might be a different kind of contamination event than in most of the data. Furthermore, the proximity to the closeby Goldcare hospital makes it possible that a contaminated car was heading for that particular point of interest, and got parked right next to a measuring mobile sensor.
Since the St. Himark Power Plant is known as a community-minded company that keeps ties with local organizations, a connection to the power plant is not unlikely. It might be possible that a company car got contaminated right at the source (the power plant) and was used in order to get to Goldcare Hospital when the radiation was picked up by the mobile sensor 29.
The sudden disappearance of the cluster is yet another indication for the theory of a moving contamination.
Another contaminated spot was recorded by mobile sensor 21 on Thursday shortly after the previously-mentioned cluster. The timeline view allows hovering to highlight the location of the peak measurement on the map, making it easy to link temporal and spatial cluster properties.
After hovering in the outlier we can see that the location of this one is the same as the big spot location. So it is part of the big cluster from where we can assume that even though this sensor stopped recording multiple times it measured reliable measurements and it is a sensor to be trusted since it measurement contamination of the same location and around the same time with mobile sensor 21. Also on Thursday just near the Powerplant we can spot some contamination in the map measured by mobile sensor 12 which is not so high.
From the contaminated spots seen on Wednesday after the earthquake and on Thursday we can say that the Downtown , Weston , Southton and West Parton are also contaminated areas since most of the contamination spots are measured there.
During Thursday and Friday the South East and the Old Town are not contaminated at all. There is an exception of southeast the ‘Wilson Forest’ were some relatively high number of contaminations were shown right after the earthquake.
On
Friday, from
4-6 am some relatively high contaminated areas were spotted in the
area of Wilson
Forest. The
measurements were done by
sensor
29.
According to social media, power plant coolant leaking might have started on 06.04.2020 10:32:00 am. After 10:32:00 we can see some more contaminated spots from the map and the contaminations rise, making the theory of contamination spreading through coolant possible.
Since the area around the power plant is obviously the most crucial in terms of possible contamination spreading, we recommend more sensors in that neighborhood. This makes it a lot easier to detect leaking radiation before it is spreaded throughout the city and cannot be controlled anymore.
In general, busy areas and points of interests like hospitals, theaters and schools should be covered by static sensors. This gives a good estimate on whether there are moving contaminations at all, while mobile sensor data can then be used for further analysis. Obviously, these have the big advantage of being location-independent and thus can capture the city’s overall contamination in a more precise way.
The trade-off for the ability to “cover more ground” is of course the vulnerability against power outages and reception losses. Even though the power grid suffered multiple hits over the whole time span, only one static sensor stopped recording, making them a lot more reliable.
Since the south-eastern part of the city is generally very uncertain in terms of the amount of data, installing more sensors in these districts would be reasonable. The size of Voronoi cells in our map view plotted around sensors and points of interest support this theory visually, since the size of a cell indicates how big of an area a single sensor is covering.
The map below shows static sensors and districts that do not have any static sensors at all. For the busy north-eastern area around the Old Town district, there is enough available mobile sensor data to counteract the lack of static sensors. However, as described before, the south-eastern region notoriously lacks data, which would make the placement of static sensors a good idea.
Furthermore, it would make sense to place static sensors on all city bridges to reliably detect whether contamination is entering or leaving the city.
4 –– Summarize the state of radiation measurements at the end of the available period. Use your novel visualizations and analysis approaches to suggest a course of action for the city. Use visual analytics to compare the static sensor network to the mobile sensor network. What are the strengths and weaknesses of each approach? How do they support each other? Limit your response to 6 images and 800 words.
Towards the end of the simulation period, we can see the south-eastern districts progressing to become very uncertain areas. It is worth noting that highest detected contaminations occur in these very sparse areas. However, as mentioned before, these areas show a high average streak length which need to be taken into consideration.
As mentioned before, static sensors are generally suitable for high-throughput areas like airports, public spaces, schools, hospitals and of course the power plant. They allow it to easily detect moving contaminations, provided they can get close enough to the measuring sensor which makes it crucial to deploy enough over the city. The more static sensors and the more evenly they are distributed, the easier the task of localizing a moving contamination will become.
Furthermore, static sensors seem to be resistant against infrastructure damages. While they probably would be affected by power outages, this seems to not have been the case in St. Himark. Since all mobile sensors except 11 constantly struggle with reception loss, the weak spot seems more to be the cell phone network affected by the earthquake which does not trouble static sensors at all.
As mentioned before, mobile sensors provide by far more spatial information about the global state of the city, while having the big trade-off as to being more vulnerable against events in their proximity.
A combination of both static sensors at busy points of interests and city borders together with a large fleet of mobile sensors would make it easy to pinpoint contaminations on the map. While their measurements might be off by a certain delta, the timestamp of peaks could help for further investigations.
5 – The data for this challenge can be analyzed either as a static collection or as a dynamic stream of data, as it would occur in a real emergency. Describe how you analyzed the data - as a static collection or a stream. How do you think this choice affected your analysis? Limit your response to 200 words and 3 images.
Our tool used the provided data as a static collection since we imagined it to be used in a post-crisis scenario for response analysis.
It is intended to analyze and understand the spread of contamination across the city and locate its potential source, allowing emergency responders to have a head-start in the next potential crisis.
Had we focused on a streaming-based approach, we would have invested more time in visualizing changes over certain timesteps and events that drastically change the city’s overall state. One example for such an event would be the drastic increase of uncertainty in south-eastern part of the city. To detect such an event in real-time would be crucial for a tool like this. It might have been necessary to have an overall score for the city’s state whereas in our chosen static approach, we were able to focus on uncertainty in individual aspects.
