Inspiration💡

Heat wave in the southern part of Europe. Greece, like other parts of southern Europe, has seen sweltering temperatures over the past week. The high temperatures — central Greece saw the mercury hit 44 degrees Celsius (111.2 Fahrenheit) — combined with high winds, have led to Greater Athens and southern Greece being put under high alert for wildfires. On July 17, 2023, wildfires hit the Greece town of Attica due to a heatwave that has been experienced in Southern Europe.

We have found that in this study area usually has fire occurrence every year, so we have tried to use models to detect burned areas for monitoring the fire event and it is good to evaluate the possible damages to land use and land cover to plan the preparation and response for the future.

What it does⚙️

The objective is to train the model for detecting the burning area in Attica, Greece from June 26 to July 14, 2023. With the result of the model, we can estimate the damage in the area and provide information for the future fire event. To achieve this objective, we utilized satellite images and open-source models to detect the burnt area by using FireCLR and SAMGeo machine learning models. And the final result is to show the burned areas in web map using Flask and Python.

How we built it🛠️

This project is built by using machine learning models integrated with geospatial analysis and visualize the results by creating the web map.

Challenges we ran into⌛

  1. While using the FireCLR model to extract the feature from our fire image, we have to crop down our image to small images and put them in a folder but the folder is limited to 10,000 images. So, our results are not very smooth because we have big cropped images with a big stride when cropped them.
  2. From the first limitations, we change the file system to save them by 1 column 1 folder but we still face the limitation of saving folders. We can’t create more than 100 folders. So, we have to crop the image with a big stride.
  3. The limitation of RAM in Colab, because of training the model using a lot of computation, the colab can crash easily.
  4. The result from FireCLR has a coarse resolution so when it is converted to polygon, it is necessary to refine or smoothen the edges.
  5. SAMGeo supports only 8-bit images, so we need to perform pre-processing on our datasets from 16-bit to 8-bit before using the model.
  6. We could only deploy our web application locally at the moment.

Accomplishments that we're proud of👏

There are 3 steps of our methods: 1. Training model with FireCLR/ Segment burned area with SAMGeo. 2. Assessment the damage from the fire event by integrating a geoinformatics system and satellite images. 3. Web map development with Python. So we integrate the knowledge in the fields of Geoinformatics, earth observation and computer science altogether in these3 phases.

What we learned✏️

  1. Utilizing a machine learning model to deal with disaster events
  2. Dealing with geospatial datatype such as vector, tiff file by using rasterio and numpy python library.
  3. Data pre-processing 4 Folium library simplifies the process of creating interactive maps in Python.
  4. Learning about unsupervised learning and contrastive learning
  5. Learning more about image change detection technique.

What's next for Inferno Tech: STEMIST HackII 2023📈

Next implementation:

  1. Improve the model to classify more fire spots in images and predict the spreading of fire to help manage and prevent disaster.
  2. Implemented with farmers to create smart farming
  3. Implemented with social media by creating the feature to report the fire by clicking on a map or using 4. Develop a web that users can send notifications to firefighters.

Built With

Share this project:

Updates