Pixel-wise feature fusion in gully susceptibility: A comparison of feed-forward neural networks and ensemble (voting, stacking) models
Vincent E. Nwazelibe, Johnson C. Agbasi, Daniel A. Ayejoto, Johnbosco C. Egbueri
Abstract
Similar to other geologic hazards, gullies pose significant challenges to Nigeria's southernmost State, requiring a reliable susceptibility mapping analysis to support decision-making. However, challenges exist regarding model recommendations, especially in studies utilizing multiple models that perform well but show visual differences, necessitating the fusion of pixel-wise features from base model predictions to present a single useable map. Although hybridised models exist, they often produce results by pairing models rather than fusing multiple model outcomes. While stacking and voting approaches could address this problem, they have remained relatively unexplored. Using a fully connected feed-forward neural network (FNN) with ensemble voting and stacking methods as pixel-wise feature fusions, this study seeks to answer these questions: How can multiple base machine learning (ML) models (Bagging, Random Forest (RF) and Extreme Gradient Boosting (XGB)) be combined? Do they improve accuracy compared to base models? How do fusion methods differ and generalize to new data? Thirteen (13) conditioning factors were used alongside complex k-fold cross-validation, training, and testing inventory structure of 574 gully and non-gully points. Our k-fold validation and testing results of the Area under the Receiver Operating Characteristic Curve (AUC) and Mean Absolute Error (MAE) show that FNN (AUC: 0.9666 & 0.9670; MAE: 0.1432 & 0.1616) strengthens the base model and generalises better than stacking (AUC: 0.9821 & 0.9598; MAE: 0.0703 & 0.1604) and voting (AUC: 0.9990 & 0.9752; MAE: 0.0532 & 0.1485) but requires parameter optimization. The study expands the knowledge about the fusion of ML methodologies within geospatial analysis and advances gully-related literature within the study area to support mitigation strategies. • Fusion models compensated for inherent weaknesses of base models, improving accuracy. • FNN fusion model, compared to stacking and voting, has more stable accuracy & error differences. • Areas at high risk were identified in southern parts alongside areas with lower risk of gullying. • Topographic and hydro-climatic factors significantly influence gully susceptibility in Anambra. • The fusion model is reliable for policymakers for natural hazards management, globally.