Comparative architectural evaluation of CNN models with alternative classifiers for UAV-based detection of damaged green zones

Unmanned aerial vehicle imagery provides a practical and high-resolution data source for monitoring the spatial condition of urban and peri-urban green zones. However, the reliable detection of damaged, stressed, or declining vegetation areas remains challenging because UAV images often contain complex background patterns, illumination variations, shadows, mixed vegetation structures, and visually similar damage symptoms. Convolutional neural networks have shown strong potential for extracting hierarchical visual features from image data, yet their conventional fully connected classification heads may not always provide the most robust decision boundary, particularly when training data are limited, class distributions are imbalanced, or subtle vegetation damage patterns must be separated. This study presents a comparative architectural evaluation of CNN-based models in which the internal components of the network, including convolutional kernel size, pooling strategy, flattening mechanism, and classification head, are systematically modified and assessed. A custom CNN architecture is used as the feature extractor, while its conventional classification head is replaced with alternative machine learning classifiers, including XGBoost, Random Forest, Support Vector Machine, and Logistic Regression. In addition, the effects of different convolutional kernels, max and average pooling operations, flatten and global average pooling representations are experimentally investigated. The intermediate outputs of convolution, pooling, and feature extraction stages are visualized to interpret how architectural modifications influence the representation of vegetation texture, color degradation, canopy gaps, and damaged regions. The proposed experimental framework enables a detailed comparison between conventional CNN classifiers and hybrid CNN–machine learning models. The study evaluates each configuration using classification metrics, visual feature analysis, and model interpretability tools. The outcome is to identify an effective CNN-based architecture for UAV imagery that improves damaged green zone detection while providing a clearer understanding of how architectural choices affect model behavior.