ML & Philanthropy — Natural Disaster Prediction & Relief Models
The average number of natural disasters in the 21st century stood at 185 (2000–2019), but 2020 has gone above that average amounting to 207 natural disasters in the first quarter of the year alone. Apart from being devastating, any form of disaster has a way of bringing people closer as they strive to help each other and fight the calamity together.
Natural disasters occur unexpectedly and are independent of human control, but thankfully, humans have evolved to a point where we now have a tool that could potentially help us predict such disasters. Data scientists all over the world agree that ML (machine learning), coupled with AI, can have philanthropic use, prevention or damage mitigation from natural disasters being one of them.
It is a given fact that we can’t really do much when it comes to prevention, but when it comes to mitigating the impact of said disasters, ML-based models are proving to be a good option. One such option is the Building Damage Detection in Satellite Imagery Using Convocational Neural Networks, where an ML approach is used to automatically process satellite data and assess the damage. This model is developed by Google and United Nations World Food Program (WFP), and the goal is to reduce the time and effort crisis workers have to put in not just to assess the damage, but also to deliver disaster aid.
Here we’ll discuss how this works and how ML has a major role to play in not just the detection model, but also other, predictive models that can be designed.
Data Science Innovative Solutions — Reducing Aid Worker Effort
The US Department of Defense (DoD) recently organized the xView2 competition where the same model was shown in action. The model works by assessing satellite images of an affected area before and after a disaster to identify damaged buildings and relaying the data to aid workers so that they can spend less time finding the affected areas.
The process is split into two core areas of feature engineering; identification of houses and then the damage done.
With the help of Google maps, two images of the affected area are taken and then analyzed by AI. In recent models, the xBD dataset is being used, which includes information about 8 types of natural disasters and 15 countries.
In any given disaster, there are at least 3 different types of damages to look at:
1. Fire damage
2. Water damage, and
3. Loss of structural integrity
The model first goes through the image of the area before the disaster to identify houses, and then to identify damage. The image is analyzed at least three times to check for each type of damage separately. Both, building identification and damage assessment, require a high-accuracy ML algorithm.
The workings of the whole algorithm are shown below:
Building identification used to be one of the biggest issues presented by the model. ML localization training took care of that. Specifically, the model used pre-trained weights, built with a VGG-16 encoder. Here is a visual representation of the architecture:
Crisis Analytics & Disaster Prediction Models
While the ML-based disaster mitigation system is in its implementation phase, the predictive models are still in the development and testing phase. Python data products are being used to do predictive analysis, while ML implementation in those models serves to enhance them.
One proposed model is a hybrid model that uses ML with Artificial Neural Networks and Data Mining approaches. Artificial Neural Networks (ANN) are algorithms in ML that have found extensive use because of the versatility they offer.
Let’s take the example of flood type disasters here. A sample time-frame is considered for the sake of this model and all the parameters are taken into account, such as weather attributes, potential human elements, monthly average rainfall, temperature, and more, to train and develop a machine learning model.
The model is built with a two-class neural network classifier that calculates the probability of floods based on those input parameters. Currently, such models in testing phases have been able to predict floods with 91.9% accuracy.
The base machine learning platform used in the study linked above is Microsoft Azure cloud-based machine and the following are the workings of the model.
In short, the basic steps involved in making a predictive model for floods can be summarized as follows:
1. Collecting the data and setting aside the training dataset
2. Training the model
3. Making monthly predictions and scoring them on the Test data
4. Evaluating the model and set it to predict future results
Although the list seems to be very simple, the work behind it is immense and susceptible to errors. This is perhaps a core critique of data science that the results often represent just the tip of the iceberg.
Why These Models Aren’t More Widespread
Both the models explained above are tailored to specific disaster types. Where the mitigation model is mostly aimed at helping aid workers in case of an urban disaster, the latter is aimed at floods. Not every disaster is as predictable as floods; earthquakes and coastal disasters, for instance.
The Bayesian Network (BN) has found extensive use in probabilistic prediction of such disasters.
A prime utilization of BN is the estimation of coastal erosion due to a rise in sea-levels, geology, and morphology. The relationship is represented as follows:
P(Ei )=∑(G,D,SLR) P(Ei│D,G,SLR,) P(D┤|G,SLR) P(G│SLR)P(SLR)
Here, the left side describes the probability of a certain amount of erosion (Ei). The right side describes the conditional probability of said erosion occurring depending on the state of coastal dunes (D), geologic setting (G), and rise in sea-level (SLR).
There is also the Regression Tree (RT) that uses linear regression to split probabilities in binary form. This leads to branching, which allows users to quickly assess the impact of any given change on the probability of disaster. This is a rather straightforward model, but the predictions are yet to be proven reliable. (Source: Evan B. Goldstein Thesis, Department of Geological Sciences, University of North Carolina)
The possibilities of philanthropy with data science are limitless, especially as more and more ML elements are incorporated into the mix, however, the main problem remains; that of it being a lengthy and sophisticated process. This sophistication is why implementation in real life is rather slow.
