Abstract
Deep learning models for flood and wildfire segmentation and object detection enable precise, real-time disaster localization when deployed on embedded drone platforms. However, in natural disaster management, the lack of transparency in their decision-making process hinders human trust required for emergency response. To address this, we present an explainability framework for understanding flood segmentation and car detection predictions on the widely used PIDNet and YOLO architectures. More specifically, we introduce a novel redistribution strategy that extends Layer-wise Relevance Propagation (LRP) explanations for sigmoid-gated element-wise fusion layers. This extension allows LRP relevances to flow through the fusion modules of PIDNet, covering the entire computation graph back to the input image. Furthermore, we apply Prototypical Concept-based Explanations (PCX) to provide both local and global explanations at the concept level, revealing which learned features drive the segmentation and detection of specific disaster semantic classes. Experiments on a publicly available flood dataset show that our framework provides reliable and interpretable explanations while maintaining near real-time inference capabilities, rendering it suitable for deployment on resource-constrained platforms, such as Unmanned Aerial Vehicles (UAVs).
Abstract (translated)
用于洪水和野火分割及目标检测的深度学习模型在嵌入式无人机平台上部署时，能够实现精准、实时的灾害定位。然而，在自然灾害管理中，其决策过程缺乏透明度，阻碍了应急响应所需的人类信任。为解决这一问题，我们提出了一个可解释性框架，用于理解广泛使用的PIDNet和YOLO架构上的洪水分割和车辆检测预测。更具体地说，我们引入了一种新颖的再分配策略，该策略扩展了逐层相关性传播（LRP）对sigmoid门控逐元素融合层的解释。这一扩展使得LRP相关性能够流经PIDNet的融合模块，覆盖整个计算图直至输入图像。此外，我们应用基于原型的概念解释（PCX）在概念层面提供局部和全局解释，揭示哪些学习特征驱动了特定灾害语义类的分割与检测。在公开洪水数据集上的实验表明，我们的框架在保持近实时推理能力的同时，提供了可靠且可解释的解释，使其适用于无人机等资源受限平台的部署。
URL
https://arxiv.org/abs/2603.23020
