Figure 6
Textual inputs include news and articles, social media, communications and press releases. Data preprocessing detects and removes non-A S C I I characters and filters languages. Data normalisation applies lowercase conversion, punctuation removal and stopword removal. Content quality control removes short texts and applies specific coding. Advanced text processing applies lemmatisation, duplicate removal and spelling correction to produce processed textual data. Visual inputs include satellite and radar data, social media, and U A V and C C T V footage. File integrity and validation check file formats and corruption. Data quality checks address resolution and clarity before quality enhancement. Content relevancy filtering removes duplicates and classifies relevance. Technical processing applies multi-spectral synthesis and cropping to produce processed visual data. Textual single-modal analytics include humanitarian needs, sentiment analysis, topic modelling and clustering, and entity recognition. Visual single-modal analytics include human and object recognition, damage and change assessment, geospatial analysis and mapless navigation. Early data fusion connects processed textual and visual data to multi-modal analytics comprising cross-modal validation, spatiotemporal analysis, multimodal learning and ensemble models with late fusion.Multimodal textual and visual data processing pipeline
Source: Authors’ own work based on Alam et al. (2019), Ullah et al. (2021), Zhang et al. (2024b)Â