EarthWise| Spot That Fire V3.0
Spot That Fire V3.0
EarthWise
Summary
WHO has defined wildfires as “An unplanned fire that burns in a natural area such as a forest, grassland, or prairie.” Most of the time, wildfires are caused as a result of natural phenomenons like lightning, extreme heat, high temperatures during summer, and drought. However, wildfires starting due to human activities are also not unheard of. Almost half of the time, reasons that cause these fires are unknown. Wildfires have impacted local ecology, animals, farmlands, regional air quality, properties and people. Prediction of wildfires thus become an important aspect in controlling these losses and also assists in providing time to take necessary safety measures.
How We Addressed This Challenge
Project:
We have named our project as "EarthWise" as we aim to incorporate machine learning intelligence to solve real world phenomenons and problems. For that purpose, we have also designed this logo to help create a better understanding of our visions and also to help in making machine learning look more visually creative. The elements in the logo indicate their uses. The hexagon at the base indicates the six steps required by machine learning models to operate properly. The planet with the tree on top indicates the field of work, which is natural earth. The shield on top indicates its use to help save the planet and to solve the problems related to natural phenomenons. The colours used are the colours that are most widely seen from space on earth.
Development:
We have developed a supervised machine learning model with comparative analysis of different algorithms for better results. A major advantage that this project has is that it can find which algorithm gives a better accuracy for the selected dataset. The algorithms that we have selected for this project are as follows:
- K Nearest Neighbours
- Support Vector Classifier
- Random Forest
- Logistic Regression
- Stochastic Gradient Descent
Goal:
Our main goal is to predict where wildfires could happen. To do that we took the areas where previous fires had taken place. Apart from that we also took the areas where fire had not spread widely. We took brightness of each region of fire and created a model where using the brightness or intensity of the area, a prediction of whether fire had escalated at that place or not could be made.
Working:
To explain the working of the project code, we have listed out the steps covered:
- Plotting histogram using brightness.
- Finding quantile of brightness, latitude and longitude.
- Dividing brightness in low, high and extreme on the basis of value and add this as a column.
- Categorising area on the basis of latitude and longitude.
- Pre-processing the data by replacing object value with integer value.
- Applying different machine learning algorithms.
Vision:
We hope that this project is able to provide a better prediction model of wildfires. We also hope that this project is able to help future works in this field.
How We Developed This Project
Inspiration:
From all the challenges that were listed, we rounded up a few which fell under our interests and aptitude. In that, we finally chose this challenge as it mainly dealt with machine learning. Machine learning being our shared interest, we also looked up for the prerequisites that this challenge required and concluded that it was possible to create an ML model that could solve the problem. We knew that machine learning is a vast field with various methodologies that could compete for a general solution.
Approach:
Our main approach to solve this problem was to identify which of the methodologies could give better results. As such, we have utilised five different machine learning on our selected dataset, and did comparative analysis.
Implementation:
To implement this, we have used open source google colab which is a jupyter notebook environment and is cloud based. We used python 3.0 as our coding language, and have utilised numpy, panda and seaborn as main python language libraries. We have also used Scikit learn, which is a free machine learning library, for implementing the various algorithms used in this project. Apart from that, we have used matlab based matplot library for plotting charts for exploratory data analysis and visualisation. We have used folium library to plot a map of wildfire data.
Setbacks and Recovery:
The main problem that we faced during this project was the type of dataset that was available in the resources. The data type of the files were either hdf5 or csv. We did not have any major experience in dealing with hdf5 file format and thus decided to take csv file. That narrowed down our field of working and we focused on the dataset that we could work with. We believe that our main contributions through this project is the comparative analysis which pits different algorithms for better accuracy on the same dataset.
How We Used Space Agency Data in This Project
From the resources given for this challenge, we have used MODIS dataset from NASA Earth Data. We have also created our own image dataset which contains images containing fire and no-fire images. For the MODIS dataset, we utilised the csv file by aptly studying the given attributes of the dataset. The data contains the details of wildfires in the north american continent. Apart from that we also visualised the attributes for a better understanding. The dataset contained details regarding brightness and area of fire. We were able to use brightness as a measure to check for prediction of fire in a particular area.
Project Demo
Data & Resources
We decided to pick the MODIS Dataset from the NASA EarthData. It is a csv file format with 13 attributes. Apart from that we have also used fire-dataset, which is a collection containing both fire and non - fire images.
The link for datasets is given below: