Machine Learning-Based Land Cover Classification and Impact Assessment in Pre-Wildfire and Post-Wildfire Areas

Focused on assessing post-wildfire vegetation changes in Mount Charleston, this study utilized Landsat 8 imagery and machine learning models. The development of the training dataset relied on US Forest Service Landscape maps. Support vector machine (SVM) and neural network (NN) models were trained using Matlab Classification Learner. The SVM with linear kernel achieved an impressive 92.5% overall accuracy, effectively delineating fire-impacted areas, revealing a 25% reduction in trees and a 22% increase in barren lands. The SVM had a mean absolute error (MAE) of 0.1264 and a root mean square error (RMSE) of 0.41 for pre-fire classification, an MAE of 0.0718 and an RMSE of 0.3812 for post-fire state. The NN model with three layers demonstrated a 96.8% overall accuracy, indicating a post-fire conversion of dense vegetation to mixed barren shrubs. The model had an MAE of 0.0141 and an RMSE of 0.1393 for the pre-fire classification maps and an MAE of 0.0054 and an RMSE of 0.1019 for the post-fire classification. The tri-layered NN performed better than its SVM in terms of overall accuracy and MAE and RMSE. The spectral signature analysis identified challenges related to misclassifications, due to classes having similar spectral signatures emphasizing the need for a finer-resolution training dataset. These findings underscore the valuable role of machine learning models in understanding and quantifying wildfire-induced ecological impacts and how these changes can alter hydrological regimes.