A Better Future - Space Apps Challenge

A Better Future| Automated Detection of Hazards

Automated Detection of Hazards

Countless phenomena such as floods, fires, and algae blooms routinely impact ecosystems, economies, and human safety. Your challenge is to use satellite data to create a machine learning model that detects a specific phenomenon and build an interface that not only displays the detected phenomenon, but also layers it alongside ancillary data to help researchers and decision-makers better understand its impacts and scope.

A hybrid machine learning model for detecting static aerosol blankets.

Summary

Hello everyone, we are team “A better future”, consists of 3 designers and 2 developers from Colombo, Srilanka. Our priority is to build a machine learning software to detect Ash plums, smoke plums and dust storms, collectively it can be considered as the aerosol blanket. The meteorological phenomenon such as dust storms can cause a considerable damage to properties and lifeforms whereas ash plums can cause a damage after a long time after occurring. Hence, we built a software to detect and identify the hazard occurring locations with its scope ,impact and to inform the relevant authorities. Features such as image save library and generating comparison report detections are included.

How We Addressed This Challenge

Our project main priority is to detect and analyze the hazard occurring locations . As the aerosol blanket that we are detecting is a collective hazard of dust storm , ash plums and smoke plums. NASA space apps 2020 challenges were provided to us to present a practical solution to serve the community by avoiding and controlling the possible threats. The aerosol blanket hazard cause considerable damage to properties and lifeforms around the world . Statically considering the damage done to the lifeform is higher than the damage done to properties. So, aerosol blanket can be considered as major threat hazard for the community. So, our machine learning model (application) detects this hazard occurring locations and its scope, impact and inform the relevant authorities.

How We Developed This Project

Machine learning model development

For this project we basically used programming languages python and C# . Python for the machine learning model and C# for the GUI design. When it comes to software modelling and for the machine learning model we have implemented python languages and its libraries such as h5, PyTorch, Pandas , TensorFlow ,Numpy and matplotlib and for preprocessing we used Keras preprocessing models and addition to that we have implemented the model hierarchy using a categorical costraintopy that is for the loss function and we have implemented several batch sizes , classes , epochs and validation splits which are useful for the machine learning model. We have implemented interpolation for several data sets such as aerosol index data set and we implemented the training data set for that model as well .In parallel we have implemented this machine learning model for other data sets as well such as sulfur dioxide and carbon monoxide data sets . By combining all together, we have collaborated these datasets and result into a hybrid model which would be our final result and for the hardware we currently using our personal laptops with a processor performance of Intel core i7 – 8550 processor and its 11 GPUs and we have extended our resources by using Microsoft azure for the better performance of the algorithms.

Machine learning model database development

Basically, we have implemented hdf5 files for our databases because they are more intuitive and more descriptive in accordance with the datasets rather than CSV or other file formats

Challenges facing during the project

We are currently facing a challenge to modelling the hybrid architecture for the machine learning model because there are several hybrid architectures that can be used but the accuracy levels are getting collaboratively degraded when we are using hybrid model and addition to that the computer resources we personally have are not sufficient for the entire project because the GPUs are not enough for running the machine learning algorithms that are related here and the GPU will definitely crash if we are going to implement the hybrid machine learning model and addition to that we are currently phasing for the collaborative detection and impact analysis . So, there are several things we have to complete for the impact analysis too.

Combining the Python machine learning model with the C# GUI to finalize the project

Since the machine learning model and the GUI is done with two different languages which are respectively Python and C# . So by using an API code we will be able to finalize the project by making a clear programming structure between the two languages

How We Used Space Agency Data in This Project

this project the datasets which we used are entirely NASA data sets provided from the link - https://earthdata.nasa.gov/earth-observation-data/near-real-time/hazards-and-disasters .As explained above in “Addressing the challenge” aerosol blanket hazard is collection of ash plums , smoke plums and dust storm. Hence to detect the aerosol blanket occurrence we used different data sets for each. (the resources and the satellite image category used is specified below)

For Ash plums we used https://earthdata.nasa.gov/earth-observation-data/near-real-time/hazards-and-disasters/ash-plumes

· Aerosol Index - (OMPS (Suomi NPP))

· Aerosol Optical Depth - (MODIS (Terra/Aqua))

· Corrected Reflectance Imagery - (MODIS Corrected Reflectance Imagery layers)

· Fire - (MODIS (Terra))

· Land Surface Reflectance - (MODIS (Terra))

· Sulfur Dioxide - (OMPS (Suomi NPP))

For smoke plums we used https://earthdata.nasa.gov/earth-observation-data/near-real-time/hazards-and-disasters/smoke-plumes

· Aerosol Index - (OMPS (Suomi NPP))

· Aerosol Optical Depth - (MODIS (Terra/Aqua))

· Corrected Reflectance Imagery - (MODIS Corrected Reflectance Imagery layers)

· Fire - (MODIS (Terra))

· Land Surface Reflectance - (MODIS (Terra))

· Sulfur Dioxide - OMPS (Suomi NPP)

For dust storms we used https://earthdata.nasa.gov/earth-observation-data/near-real-time/hazards-and-disasters .

· Aerosol Optical Depth - (MODIS (Terra/Aqua))

· Corrected Reflectance Imagery - (MODIS Corrected Reflectance Imagery layers)

· Dust - AIRS (Aqua) – Dust score (Day/Ocean)

· Land Surface Reflectance - (MODIS (Terra))

So we basically implemented hdf5 file formats for the training of machine learning model . For that first we have classified file formats in regarding NASA space agency data. Then we identify the propositional file type as the hdf5 file in which the data is organized in a hierarchical manner. Afterwards we have implemented several preprocessing algorithms on the data sets that we have selected.

Project Demo

The image above shown is the GUI prototype of the machine learning model we made. It includes many functions as a menu option , search option to search specific satellite images from the library , Library option to save past satellite images in case if we want to get them easily , save button to save the satellite image in the library after the detection analysis is completed and most importantly Report generating button to generate reports which are comparison reports with the current satellite image using for detection or previously library saved satellite images (two satellite images compared). Additionally there are 5 buttons in the top right side corner . they are respectively scope button , impact button , snapshot button , about button, and help button (some of them are included in the menu dropdown menu too for the convenience of the user) . Specifically considering scope function and impact function two of the most important functions in this model .It allows to determine the scope and impact severity of the hazard (on that location)

The above shown image is the form design of the Library . It is built to save previous date satellite images if the user wants to view detection or generate report (individually or comparison) of the past days . Also we have provided and upload button to upload past satellite images to the library.