Approach

Our state of the art model follows a two-fold multi-modal approach that includes:

• Temporal Prediction of the occurrence of wildfire at hotspot locations within the next 2 hours anywhere in United States
• Spatial Detection of the presence of fire in digital camera imagery

• For the early detection, we have implemented an many-many RNN model that will take previous data as input and gives the output as future latitude and longitude value of the satellite and the confidence value of catching wildfire on that position. The RNN model consists of 2 LSTM cells with 32 and 16 hidden layers in each LSTM cell respectively. 
• The RNN model is trained on weekly data from MODIS C6 (Aqua) Active Fire Data Repository

• For detection of fire from digital image, we have used a Convolutional Neural Network model (Inception-v3). This model is trained on labeled digital image, and we have used data augmentation techniques to increase the accuracy of the training process. Given a satellite image, the model can detect with high confidence if the image contains fire or not. This model is inspired from Deep Residual Learning model by Kaiming H. et al.

Dataset

- Our Temporal model has been extensively trained on 8806 temporal records from the MODIS C6 Active Fire Data spanning from 09/26/2020 to 10/03/2020 with following feature-set.

- Our Spatial model has been trained on 2000 labeled street images with fire and neutral (non-fire) visuals acquired from DeepQuest AI image repositories.

RESULTS

# RNN Results

MSE of the RNN model on train set is 0.0206 and on validation set is 0.0211, as we normalized the confidence value from 0 to 1.

#CNN Results

Our CNN model achieves 94% validation accuracy with only 14.5% validation loss.

Meet The People Behind The Magic

Sahara Ali is pursuing PhD in Information Systems at UMBC. She works on Applied Machine Learning for Earth Science at the Big Data Analytics Lab. Apart from scribbling code and spoiling logics, she sometimes bakes cupcakes too.

Sahara Ali

Masud Ahmed

Masud Ahmed believes in enjoying every moment of life, as life is short but beautiful. He is currently pursuing a Ph.D. in information system. His primary research work is focused on computer vision and machine learning. Besides, he is also interested in quantum physics

Yue Jeanette Huang is a current PhD student in Information System at UMBC. She is now working on applying the Byzantine Fault Tolerant protocol in the Power grid system. She is totally a cat-lover.

Yue Jeanette Huang

Sabrina Mamtaz Nourin is graduate student at UMBC. She is pursuing Ph.D in Information Systems. She is working on Software Security based on NLP and Semantics. She is a bookaholic, and you can kidnap her with a good book and some foods!

Sabrina Mamtaz Nourin

Rishabh Balse is pursuing his BTech in Computer Engineering from India. As an aspiring Android developer, Rishabh wishes to take on projects that helps the community in one way or another. Recommend him a good movie or an anime to make his day :)

Rishabh Balse

Faizan Elahi is UX Designer. He likes to solve complex problems. He has a BS in Computer Sciences from University of Engineering and Technology Lahore. Curiosity is the only thing that drives him.

Faizan Elahi