Pixel at NakbaArchiveClassifier Shared Task: ConvNeXt-Based Ensemble for Destruction Detection

Abstract

This paper describes our submission to the Nakba Image Classification Shared Task at the Nakba-NLP 2026 workshop. The task requires binary classification of social media images into two categories: destruction and not_destruction. The dataset includes approximately 1,600 annotated development images and 400 held-out test images, collected from Instagram posts published in Gaza between October 2023 and December 2025. High variability in viewpoint, lighting, and image quality, coupled with the inherent complexities of identifying structural damage in dense urban environments, makes this task particularly challenging. Our system utilizes a pretrained ConvNeXt-Tiny backbone fine-tuned through a stratified 5-fold cross-validation framework. To mitigate class imbalance, we implement a weighted cross-entropy loss function. During the inference phase, we employ an ensemble strategy that averages predictions across all five fold-specific models, and test-time augmentation (TTA) is applied to enhance robustness. The final ensemble achieved a Macro F1-score of 0.8952 and an accuracy of 0.9055 on the official test set. Our results suggest that the integration of modern convolutional architectures with robust ensembling and augmentation strategies provides a reliable baseline for automated destruction detection.

Resources

Details