AmMoPy/Wildfire_Prediction_Challenge

Predicting wildfire burned area in different locations across Zimbabwe over the period from 2014 till 2017 # Overview and objectives: - Create a machine-learning model capable of predicting the burned area in different locations across Zimbabwe over 2014 to 2017. - Predictions are the proportion of the burned area per area square, with values between 0 and 1. - There are 533 area squares each with a unique ID ranging from 0 to 532 - Training data is aggregated on burned areas across Zimbabwe for each month since 2001 up to the end of 2013, there are 156 record(13 years * 12 months) for 533 areas for a total of ~83k records (156 * 533) This is a competition notebook, check Leaderboard # Evaluation: - Root Mean Squared Error `RMSE`. # Modeling and Feature Engineering: - Overall objective and expected model behavior: - My goal is to build a model that captures the signals from repetitive temporal, interactions and proximity features. - Modeling Overview: - Running progressive 4 fold validation, re-training the model with each year addition and feature engineering within fold. `RMSE` will be calculated per fold and final evaluation will be based on Last fold score as it is most representative given full training data. Its expected to have a decreasing score per fold. - Modeling steps - per fold: - Feature Engineering: - This includes experimenting with different combination of features and models, Experimenting included: - Temporal features: given the fact that fire incidents peaks and repeats during specific months - Features that capture change over time: this has been tested across all zones and within zones during the full training period and year on year. Examples: rolling mean, std, var, cov and z-scores using different rolling windows specially those near peak months. - Feature interaction, ranges and ratios: for example, elevation multiply wind speed, Vapor pressure deficit divided Vapor pressure, max temperature - min temperature and so on. - Feature grouping: grouping features then calculating dispersion and interaction statistics within groups. Grouping is d …