Debris detection using a U-Net model trained on post-event data from extreme hydrological disasters in southern Brazil in 2023 and 2024.

This repository provides the deep learning component of a debris detection workflow based on a U-Net convolutional neural network (CNN) trained on very high-resolution post-disaster orthomosaics acquired in southern Brazil after extreme hydrological events in 2023 and 2024. The model was developed for pixel-wise semantic segmentation of two classes: debris and background. Training was performed in ArcGIS Pro 3.6.1 using the Image Analyst extension and the Train Deep Learning Model tool. The training dataset was built from manually interpreted samples extracted from orthomosaics with ground sampling distance (GSD) ranging from approximately 3.10 to 3.25 cm/pixel, enabling detailed representation of complex urban post-flood surfaces. A total of 213 labeled samples were used, including 132 debris polygons (23,501 m²) and 81 background polygons (13,468 m²). These samples generated 6,000 training chips with fixed dimensions of 256 × 256 pixels. Model calibration followed an iterative, controlled, and traceable experimentation process, in which parameter combinations were tested individually under the same validation protocol and retained only when they produced consistent improvements in the target-class metrics, especially precision, recall, and F1-score for the debris class. Different encoder backbones were evaluated during the optimization process. Initial experiments with ResNet-50 yielded limited performance for debris detection, with F1 values remaining near 0.40. The transition to ResNet-101 represented the main structural improvement in the training trajectory, substantially increasing the model’s ability to recover debris patterns in heterogeneous post-disaster scenes. To improve class discrimination under challenging visual conditions, the training process incorporated focal loss, aiming to reduce the dominance of easy examples and increase sensitivity to ambiguous regions where debris visually competes with mud, shadows, exposed soil, and degraded urban structures. The loss configuration was further refined with Dice loss fraction values tested at 0.5, 0.6, and 0.7, in order to improve overlap between predicted and reference masks and balance the learning process between classes. The influence of mixup regularization was also evaluated, but its deactivation produced better results under the ResNet-101 configuration, likely because it preserved object boundaries more effectively and reduced excessive smoothing of class transitions. A batch size of 16 was adopted as a balance between numerical stability and computational feasibility for 256 × 256 image chips. In addition, 20% of the samples were reserved for validation. The number of training epochs was also empirically adjusted. While shorter runs did not allow the model to fully consolidate feature learning under the ResNet-101 backbone, extending the training to 70 epochs produced the most stable and accurate results. Additional refinements beyond this point did not improve performance and instead reintroduced instability. The calibration trajectory comprised 13 experiments, each representing a distinct parameter configuration assessed according to the F1-score of the debris class. The final selected model reached an operational performance threshold above 0.70 F1-score for debris detection and showed stable behavior in repeated tests, indicating that the trained U-Net is suitable for extracting debris masks from very high-resolution post-flood orthomosaics in complex urban environments. This repository is intended to support reproducibility, methodological transparency, and future adaptations of CNN-based debris mapping in disaster-related remote sensing applications. Evaluation Metric Value Precision_fundo 0.873182 Precision_detrito 0.753144 Recall_fundo 0.870673 Recall_detrito 0.808421 F1_fundo 0.871926 F1_detrito 0.779804