Improved ConvNeXt for Emergency Image Classification

The timeliness of natural disaster response is crucial for minimizing losses. Traditional satellite imagery-based recognition is limited by cost, environmental occlusions, and resolution, making it difficult to meet the demands of emergency scenarios. Unmanned Aerial Vehicle (UAV) technology, with its advantages of high mobility and a wide field of view, has become an important method for disaster image acquisition. This paper addresses the class imbalance problem in the AIDER (Aerial Image Database for Emergency Response) dataset, which contains five categories: collapsed buildings, fire, flood, normal scenes, and traffic accidents. During the experiment, we first performed data augmentation on the dataset images. Then, we trained various models on the dataset and selected the ConvNeXt model for this study after careful consideration. After selecting the model, we experimented with inserting different attention mechanism modules at various positions in the model to see if we could improve the accuracy. After multiple experiments, we finally proposed this improved ConvNeXt-Tiny model for disaster image classification. The improved model embeds a Squeeze-and-Excitation (SE) attention module into stage 3 of ConvNeXt-Tiny (the mid-to-high-level feature extraction stage) to enhance key semantic information through channel-wise feature recalibration. The classification head is also adjusted to fit the 5-class task. Experimental results show that, compared to the original ConvNeXt-Tiny and other models like MobileNetV3 and Xception, the improved model achieves a classification accuracy of $\mathbf{9 7. 2 8 \%}$ on the AIDER dataset, significantly enhancing its ability to recognize disaster features. Heatmap analysis reveals that the SE module effectively guides the model to focus on key image areas (such as the center of collapsed buildings and flood boundaries), reducing misjudgments caused by variations in resolution and lighting. This study provides a lightweight and high-precision image classification solution for UAV-based disaster response systems.