Business Problem

Firefighting resource is crucial in containing wildfires and should be allocated correctly when needed. MODIS historical fire pixel data contains useful spatial information that can be used to identify patterns over time. Meteorological indicators can also provide information regarding weather conditions that causes wildfire. Soil quality indicators give information regarding drought possibilities. Though, MODIS can identify large fires with higher accuracy, it does have too many false alarms. It is expensive to use resources for false alarms. Thus, accurate prediction of fire is crucial to minimize the cost and for early preparation.

For this project, I would like to answer following questions: Do weather and soil conditions has any significance in wildfires? Can we identify true fires more accurately by combining weather and soil data to fire pixels data? Can we identify the fire prone area before fire happens, using historical fire pixels data combined with weather indicators and soil quality indicator and location information?

Techniques

     

For the second part of the problem, I used classification model of Random Forest and Support Vector Machines. I implemented model selection and evaluation step by comparing precision and accuracy rate of both model and further tuned hyperparameters to increase accuracy. Random Forest outperformed with above 98% accuracy in identifying true fire pixels.

For the last part, I attempted to use time series classification. In this I transformed the data into 3-Dimensional and implemented LSTM model for classification. Model accuracy was 75% and precision was .50.

Future Applications

Project Duration

Key Skills