|
چکیده
|
Given the inherent limitations in the results of modeling due to the need for trial and error or repeated execution with various combinations using a small number of balanced training datasets, the necessity of utilizing an imbalanced dataset to address this challenge appears to be essential. We trained and compared a hybrid machine learning approach using a Rotation Forest-based Alternating Decision Tree (RFADT) with a deep learning Convolutional Neural Network (CNN) models on both balanced (1n; n is the number of flood events) and imbalanced (2n, 3n, 4n, 5n, 6n, 7n, 8n, 9n, and 10n) datasets to produce urban flood susceptibility maps in Sanandaj City, Iran. The models were built and validated using 174 flood locations and 19 conditioning factors. We evaluated model performance using several statistical metrics, including sensitivity, specificity, accuracy, Precision-Recall (PR) curves, MAE, RMSE, and the Area Under the Curve (AUC). The results showed that RF and CNN outperformed the ADT in all cases except in the 9n dataset. More importantly, the ADT, RFADT, and CNN models exhibited a significant and consistent increase in AUC (indicating better prediction) as the size of the imbalanced dataset increased from 1n to 10n. Results of the validation dataset during modeling processing showed that: the ADT model performed best with an imbalanced dataset of 10n (MAE = 0.215; RMSE = 0.282; AUC = 0.850); the RFADT model performed best with an imbalanced dataset of 8n (MAE = 0.216; RMSE = 0.282; AUC = 0.912) and 6n (MAE = 0.261; RMSE = 0.317; AUC = 0.884); and CNN model performed best with an imbalanced dataset of 8n (MAE = 0.102; RMSE = 0.319; AUC = 0.925) and 10n (MAE = 0.086; RMSE = 0.293; AUC = 0.898). These findings also suggest that imbalanced datasets (specifically 6n, 8n, and 10n) are more effective in reliable urban flood susceptibility mapping than a balanced dataset (AUCADT = 0.774). The use of imbalanced datasets of 6n, 8n, and 10n can be a valuable approach in cases where balanced datasets underperform.
|