Hope at NakbaArchiveClassifier Shared Task: Transfer Learning-Based CNN Models for Infrastructure Damage Detection

Abstract

This paper describes Team Hope’s system for the NakbaArchiveClassifier Shared Task at Nakba-NLP 2026. The task focuses on binary classification of social media images into two categories: destruction and not_destruction. We evaluated multiple convolutional neural network architectures using transfer learning, including ResNet34, ResNet50, EfficientNet-B0, and a fine-tuned ResNet34 variant with staged training. All models were initialized with ImageNet pretrained weights and fine-tuned on the provided dataset of 2,001 images. The dataset is moderately imbalanced and contains visually diverse Instagram images depicting intact and damaged infrastructure. Our best-performing model, ResNet34 trained for 25 epochs with Adam optimizer and a learning rate of 1e-4, achieved 81% accuracy on the evaluation platform. We provide a comparative analysis of the tested architectures and discuss the impact of model depth, training duration, and class imbalance. Given the political and ethical sensitivity of the dataset, we also include a discussion of responsible AI considerations and potential limitations. Our findings suggest that moderate-depth architectures can generalize effectively in low-resource, contextually complex visual classification tasks.

Resources

Details