Livoa LogoLivoa

Data Acquisition and Inventory Development

Data Acquisition Sources

Active Fire Inventory

Predictor Variables

NASA FIRMS VIIRS, MODIS (2012–2024, 13 years)

Topographic (SRTM DEM, GEE): Slope, Aspect, Elevation, TWI

Spectral/Biophysical (Landsat 8 GEE): NDVI, NDWI, NBR

Climatic: Precipitation, LST

Land use Land cover

Anthropogenic (OSM): Distance to Roads, Distance to Settlements

Raw Data Layers

Feature Engineering and Dataset Balancing

Train/Test/Validation Data Split

Final Predictor Variables
Deep Neural Network (DNN) • Architecture: Sequential MLP (Input, Hidden 64/32 ReLU, Sigmoid Output) • Regularization: Dropout (0.3–0.4), L2 penalties
Random Forest (RF) • Hyperparameter Tuning: 5-Fold CV (mtry, nodesize, ntree) • Final Config: 700 trees, mtry = 2
Trained Models

Predictive Modeling and Algorithmic Execution

Spatial Pre-processing and Standardization

Multicollinearity Assessment (VIF) (VIF < 10)

Resampling

Bilinear/Nearest Neighbor

Absence Sampling Random non-fire points generation

Reference Alignment

UTM Projection

Delineation Masking and clipping to Madhesh Province Boundary

Final Forest Fire Susceptibility Mapping

Spatial Thinning

Standardize Spatial Data

Training Data (70%)

Performance Metrics • Confusion Matrices • ROC Curves: Accuracy, AUC, Kappa, Recall

Testing Data (30%)

Validation and Spatial Synthesis

Susceptibility Mapping (terra::predict function)

Risk Zonation 5 equal-interval levels: Very Low, Low, Moderate, High, Very High

Final Forest Fire Susceptibility Map of Madhesh Province

(RF/ DNN)

Areal Statistics (Percentage & Sq. Km Coverage)

Validation Data (15%)

FF

by KD

0
0 uses