Textual Modalities Research Articles

Military Image Captioning for Low-Altitude UAV or UGV Perspectives

Low-altitude unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs), which boast high-resolution imaging and agile maneuvering capabilities, are widely utilized in military scenarios and generate a vast amount of image data that can be leveraged for textual intelligence generation to support military decision making. Military image captioning (MilitIC), as a visual-language learning task, provides innovative solutions for military image understanding and intelligence generation. However, the scarcity of military image datasets hinders the advancement of MilitIC methods, especially those based on deep learning. To overcome this limitation, we introduce an open-access benchmark dataset, which was termed the Military Objects in Real Combat (MOCO) dataset. It features real combat images captured from the perspective of low-altitude UAVs or UGVs, along with a comprehensive set of captions. Furthermore, we propose a novel encoder–augmentation–decoder image-captioning architecture with a map augmentation embedding (MAE) mechanism, MAE-MilitIC, which leverages both image and text modalities as a guiding prefix for caption generation and bridges the semantic gap between visual and textual data. The MAE mechanism maps both image and text embeddings onto a semantic subspace constructed by relevant military prompts, and augments the military semantics of the image embeddings with attribute-explicit text embeddings. Finally, we demonstrate through extensive experiments that MAE-MilitIC surpasses existing models in performance on two challenging datasets, which provides strong support for intelligence warfare based on military UAVs and UGVs.

CLIP-FSSC: A transferable visual model for fish and shrimp species classification based on natural language supervision

Fish and shrimp species classification is a practical need in the field of aquaculture. Traditional classification method includes extracting modal features of the single image and training them on downstream datasets. However, this method has the disadvantage of requiring manual annotated image data and significant training time. To address these issues, this paper introduces a method named CLIP-FSSC (Contrastive Language–Image Pre-training for Fish and Shrimp Species Classification) for zero-shot prediction using a pre-trained model. The proposed method aims to classify fish and shrimp species in the field of aquaculture using a multimodal pre-trained model that utilizes semantic text description as an image supervision signal for transfer learning. In the downstream fish dataset, we use natural language labels for three types of fish - grass carp, common carp, and silver carp. We extract text category features using a transformer and compare the results obtained from three different CLIP-based backbones for the image modality - vision transformer, Resnet50, and Resnet101. We compare the performance of these models with previous methods that performed well. After performing zero-shot predictions on samples of the three types of fish, we achieve similar or even better classification accuracy than models trained on downstream fish datasets. Our experiment results show an accuracy of 98.77 %, and no new training process is required. This proves that using the semantic text modality as the label for the image modality can effectively classify fish species. To demonstrate the effectiveness of this method on other species in the field of aquaculture, we collected two sets of shrimp data - prawn and cambarus. Through zero-shot prediction, we achieve the highest classification accuracy of 92.00 % for these two types of shrimp datasets. Overall, our results demonstrate that using a multimodal pre-trained model with semantic text description as an image supervision signal for transfer learning can effectively classify fish and shrimp species with high accuracy, while reducing the need for manual annotation and training time.

A multi-task framework based on decomposition for multimodal named entity recognition

Given a text-image pair, Multimodal Named Entity Recognition (MNER) is the task of identifying and categorizing entities in the text. Most existing work performs named entity labeling directly using final token representations derived by fusing image and text representations. Although they achieve promising results, these work may fail to effectively exploit text and image modalities. This is because they neglect the difference in the role of the two modalities: text modality can detect the boundary of an entity, while image modality is introduced to disambiguate the category of the entity. Based on these findings, in this paper, we construct two auxiliary tasks based on the decomposition strategy and propose a multi-task framework for MNER. Specifically, we first decompose MNER into two auxiliary tasks: entity boundary detection task and entity category classification task. Here, the former treats only the text modality as input and outputs the boundary labels, since it can achieve satisfactory boundary results by itself. The latter uses two modalities to yield category labels where image modality is dedicated to disambiguating categories. These two auxiliary tasks allow the effective exploitation of text and image modalities and put them back into their respective roles. Then, we vectorize their results to improve entity recognition using label clues from auxiliary tasks. Finally, we fuse features from text and image modalities and label embeddings from auxiliary tasks to fulfill MNER. Experimental results on two widely used MNER datasets show that our framework can yield new SOTA performance.

Deep multi-task learning based detection of correlated mental disorders using audio modality

The existence of correlation among mental disorders is a well-known phenomenon. Multi-task learning (MTL) has been reported to yield enhanced detection performance of a targeted mental disorder by leveraging its correlation with other related mental disorders, mainly in textual and visual modalities. The validation of the same on audio modality is yet to be explored. In this study, we explore homogeneous and heterogeneous MTL paradigms for detecting two correlated mental disorders, namely major depressive disorder (MDD) and post-traumatic stress disorder (PTSD), on a publicly available audio dataset. The detection of both disorders is interchangeably employed as an auxiliary task when the other is the main task. In addition, a few other tasks are employed as auxiliary tasks. The results show that both MTL paradigms, implemented using two considered deep-learning models, outperformed the corresponding single-task learning (STL). The best relative improvement in the detection performance of MDD and PTSD is found to be 29.9% and 28.8%, respectively. Furthermore, we analyzed the cross-corpus generalization of MTL using two distinct datasets that involve MDD/PTSD instances. The results indicate that the generalizability of MTL is significantly superior to that of STL. The best relative increment in the cross-corpus generalization performance of MDD and PTSD detection is found to be 25.0% and 56.5%, respectively.

Deep learning for cross-domain data fusion in urban computing: Taxonomy, advances, and outlook

As cities continue to burgeon, Urban Computing emerges as a pivotal discipline for sustainable development by harnessing the power of cross-domain data fusion from diverse sources (e.g., geographical, traffic, social media, and environmental data) and modalities (e.g., spatio-temporal, visual, and textual modalities). Recently, we are witnessing a rising trend that utilizes various deep-learning methods to facilitate cross-domain data fusion in smart cities. To this end, we propose the first survey that systematically reviews the latest advancements in deep learning-based data fusion methods tailored for urban computing. Specifically, we first delve into data perspective to comprehend the role of each modality and data source. Secondly, we classify the methodology into four primary categories: feature-based, alignment-based, contrast-based, and generation-based fusion methods. Thirdly, we further categorize multi-modal urban applications into seven types: urban planning, transportation, economy, public safety, society, environment, and energy. Compared with previous surveys, we focus more on the synergy of deep learning methods with urban computing applications. Furthermore, we shed light on the interplay between Large Language Models (LLMs) and urban computing, postulating future research directions that could revolutionize the field. We firmly believe that the taxonomy, progress, and prospects delineated in our survey stand poised to significantly enrich the research community. The summary of the comprehensive and up-to-date paper list can be found at https://github.com/yoshall/Awesome-Multimodal-Urban-Computing.

Hypergraph clustering based multi-label cross-modal retrieval

Most existing cross-modal retrieval methods face challenges in establishing semantic connections between different modalities due to inherent heterogeneity among them. To establish semantic connections between different modalities and align relevant semantic features across modalities, so as to fully capture important information within the same modality, this paper considers the superiority of hypergraph in representing higher-order relationships, and proposes an image-text retrieval method based on hypergraph clustering. Specifically, we construct hypergraphs to capture feature relationships within image and text modalities, as well as between image and text. This allows us to effectively model complex relationships between features of different modalities and explore the semantic connectivity within and across modalities. To compensate for potential semantic feature loss during the construction of the hypergraph neural network, we design a weight-adaptive coarse and fine-grained feature fusion module for semantic supplementation. Comprehensive experimental results on three common datasets demonstrate the effectiveness of the proposed method.

A Multimodal Discourse Analysis of Advertisements and Movie Posters

Multimodal discourse analysis, grounded in Halliday's Systemic Functional Grammar, has been widely applied in advertising, movie posters, and e-commerce. Studies often use Kress and Van Leeuwen's visual grammar to understand how images convey meaning. It will compare the multimodal theoretical analysis of movie posters and ads, focusing on textual, color, and layout modalities, as well as stylistic features and internal modalities. Additionally, we will explore enhancing user experience in advertisements through multimodal analysis, aiming to increase impact, expressiveness, user engagement, and precise placement.

Multimodal semantic enhanced representation network for micro-video event detection

Currently, micro-videos have gained widespread acceptance as a prominent form of user-generated content across various social media platforms. Accurate event analysis of micro-videos can greatly enhance the many diverse social media platforms applications. Although some studies have shown promising results from a multimodal perspective, there is still a challenge in extracting informative cues from inaccurate modalities, particularly for text modality that is prone to inaccuracies and noise. In this paper, we propose a multimodal semantically enhanced representation network (MSERN) for micro-video event detection. To better address inaccurate and noisy text sentences, we first extract visual concepts in the form of adjective-noun pairs (ANPs), through a fine-grained common representation module, to complement the textual descriptions. To maximize the acquisition of modality-specific cues from both visual and textual modalities, we then implement a coarse-grained private representation module to ensure that private representations encompass unique facets of the modalities beyond the common perspective. Finally, because two modules will collaborate, the fine-grained common and coarse-grained private representations are integrated to ensure a reinforced micro-video representation. We evaluate our proposed method on a micro-video event detection dataset and the experimental results show a superior performance compared to the state-of-the-art methods.

Auxiliary audio–textual modalities for better action recognition on vision-specific annotated videos

Most current audio–visual datasets are class-relevant, where audio and visual modalities are annotated. Thus, current audio–visual recognition methods apply cross-modality attention or modality fusion. However, leveraging the audio modality effectively in vision-specific videos for human activity recognition is of particular challenge. We address this challenge by proposing a novel audio–visual recognition framework that effectively leverages audio modality in any vision-specific annotated dataset. The proposed framework employs language models (e.g., GPT-3, CPT-text, BERT) for building a semantic audio–video label dictionary (SAVLD) that serves as a bridge between audio and video datasets by mapping each video label to its most K-relevant audio labels. Then, SAVLD along with a pre-trained audio multi-label model are used to estimate the audio–visual modality relevance. Accordingly, we propose a novel learnable irrelevant modality dropout (IMD) to completely drop the irrelevant audio modality and fuse only the relevant modalities. Finally, for the efficiency of the proposed multimodal framework, we present an efficient two-stream video Transformer to process the visual modalities (i.e., RGB frames and optical flow). The final predictions are re-ranked with GPT-3 recommendations of the human activity classes. GPT-3 provides high-level recommendations using the labels of the detected visual objects and the audio predictions of the input video. Our framework demonstrated a remarkable performance on the vision-specific annotated datasets Kinetics400 and UCF-101 by outperforming most relevant human activity recognition methods.

TCHFN: Multimodal sentiment analysis based on Text-Centric Hierarchical Fusion Network

Multimodal sentiment analysis (MSA) has become a popular field of research in recent years. The aim is to combine the three modalities of text, video, and audio to obtain comprehensive emotional information. However, current research often treats these three modalities equally, downplaying the crucial role of text modality in MSA and ignoring the redundant information generated during multimodal fusion. To address these problems, we propose the Text-Centric Hierarchical Fusion Network (TCHFN), employing a hierarchical fusion strategy. In this framework, low-level fusion involves cross-modal interactions between pairs of modalities, while high-level fusion extends these interactions to involve all three modalities. Through the design of the Cross-modal Reinforced Transformer (CRT), we achieve cross-modal enhancement of the target modality, facilitating a nuanced fusion process with text serving as its core. Additionally, we design Text-Centric Contrastive Learning (TCCL) to align non-text modalities with the text modality, emphasising the central role of text in the fusion process. After fusion, a multimodal fusion output gate is employed to mitigate redundant information within the multimodal fusion representation, which is subsequently processed by a linear layer for prediction. Simultaneously, to fully leverage limited labelled datasets, we introduced knowledge distillation. This approach involves preserving the model parameters that yield the best performance during training as a teacher model. The teacher model aids in capturing rich emotional information, enabling the model to transcend local optima and discover more optimal parameters, thereby enhancing overall model performance. Extensive experiments on the CMU-MOSI, CMU-MOSEI, and CH-SIMS datasets demonstrate the superiority of our model over state-of-the-art methods in MSA tasks.

ESC-ZSAR: Expanded Semantics from Categories with Cross-Attention for Zero-Shot Action Recognition

Zero-shot action recognition endeavors to identify novel action categories not encountered during training by aligning a joint semantic space. However, despite advancements, zero-shot action recognition still needs to grapple with the inadequate semantic representation of seen data, hindering the transfer of diverse action videos. This study introduces a novel framework combining video, optical flow, and expanded label description via a cross-attention mechanism. This integration facilitates the capture of low and high-level motion dynamics, effectively bridging the domain gap between the video and text modalities. The proposed approach of generating expanded label descriptions efficiently enhances semantic information, thus ameliorating zero-shot transferability and providing a comprehensive grasp of semantics and motion. The temporal shuffle and alignment module is designed to enhance the generalization ability of image sequences by capturing discriminative high-level motions through frame sorting. The efficacy of the proposed method is validated through extensive experiments on three benchmark datasets, namely Kinetic-600, UCF-101, and HMDB-51. Notably, our model achieves state-of-the-art results in the zero-shot action recognition task.

Multi-task disagreement-reducing multimodal sentiment fusion network

Existing multimodal sentiment analysis models can effectively capture sentimental commonalities between different modalities and possess high sentimental acquisition capability. However, there are still shortcomings in the model's analysis and recognition abilities when dealing with samples that exhibit sentimental polarity disagreement between different modalities. Additionally, the dominance of the text modality in multimodal models, particularly those pre-trained with BERT, can hinder the learning of other modalities due to its richer semantic information. This issue becomes particularly pronounced in cases where there is a conflict between multimodal and textual sentimental polarities, often leading to suboptimal analytical results. Besides, the classification ability of each modality is also suppressed by single-task learning. In this paper, We propose a Multi-Task disagreement-Reducing Multimodal Sentiment Fusion Network (MtDr-MSF), designed to enhance the semantic information of non-text modalities and reduce the dominant impact of the textual modality on the model, and to improve the learning capabilities of unimodal networks. We conducted experiments on multimodal sentiment analysis datasets, CMU-MOSI, CMU-MOSEI, and CH-SIMS. The results show that our method outperforms the current SOTA method.

Sentiment analysis of social media comments based on multimodal attention fusion network

Social media comments are no longer in a single textual modality, but heterogeneous data in multiple modalities, such as vision, sound, and text, which is why multimodal sentiment analysis strategies has been introduced. However, among the multimodal sentiment analysis domains, a majority of the current multimodal sentiment analysis models employ the Transformer architecture due to its great impact and benefits, thereby leading to an augmentation in resource overhead. In this paper, a multimodal attention fusion (MAF) network model is proposed for sentiment analysis of multimodal data. MAF is mainly composed of the cross attention and residual unit. The Cross Attention Unit is designed to select one core modality out of three modes, while the remaining modes serve as base modalities. The core modality is then combined with the base modality information to facilitate significant interaction between the two modalities, resulting in three sets of two-by-two attention computations. Moreover, a residual unit is employed to integrate the overall information into the attentional information. This approach not only enables modality-to-modality interaction, but also supplements the overall information. In the end, experiments are conducted on two publicly available multimodal sentiment analysis datasets from Carnegie Mellon University(CMU), CMU-MOSEI (abbreviated as MOSEI) and CMU-MOSI (abbreviated as MOSI), to validate that the method achieves high performance while removing the complex structure, and is comparable to the State-Of-The-Art(SOTA) model with high-performance A100 and V100 Graphics Processing Units(GPU) in an ordinary hardware environment.

On the Opportunities and Challenges of Foundation Models for GeoAI (Vision Paper)

Large pre-trained models, also known as foundation models (FMs), are trained in a task-agnostic manner on large-scale data and can be adapted to a wide range of downstream tasks by fine-tuning, few-shot, or even zero-shot learning. Despite their successes in language and vision tasks, we have not yet seen an attempt to develop foundation models for geospatial artificial intelligence (GeoAI). In this work, we explore the promises and challenges of developing multimodal foundation models for GeoAI. We first investigate the potential of many existing FMs by testing their performances on seven tasks across multiple geospatial domains, including Geospatial Semantics, Health Geography, Urban Geography, and Remote Sensing. Our results indicate that on several geospatial tasks that only involve text modality, such as toponym recognition, location description recognition, and US state-level/county-level dementia time series forecasting, the task-agnostic large learning models (LLMs) can outperform task-specific fully supervised models in a zero-shot or few-shot learning setting. However, on other geospatial tasks, especially tasks that involve multiple data modalities (e.g., POI-based urban function classification, street view image–based urban noise intensity classification, and remote sensing image scene classification), existing FMs still underperform task-specific models. Based on these observations, we propose that one of the major challenges of developing an FM for GeoAI is to address the multimodal nature of geospatial tasks. After discussing the distinct challenges of each geospatial data modality, we suggest the possibility of a multimodal FM that can reason over various types of geospatial data through geospatial alignments. We conclude this article by discussing the unique risks and challenges to developing such a model for GeoAI.

Cross-modal semantic aligning and neighbor-aware completing for robust text–image person retrieval

Most existing text–image person re-identification (TIReID) methods are performed in an ideal environment where both image and text instances are fully intact and identity annotated. However, in real-world open environments, these ideal assumptions often cannot be satisfied. In this study, we are the first to explore how to enhance the robustness of the TIReID model to better adapt to open environments, focusing on two aspects: 1) unlabeled multi-modal data, and 2) incomplete multi-modal data. Here, unlabeled multi-modal data faces two key issues: intra-class variations and cross-modal aligning problems, while incomplete data leads to data semantic inconsistency problems and severe performance degradation. In order to tackle the above issues, we propose a novel Cross-modal Semantic Aligning and Neighbor-aware Completing (CANC) method for robust text–image person re-identification (RTIReID). Specifically, to tackle the intra-class variation problem, we propose intra-view prototype contrastive matching for image and text modalities, which helps the model learn intra-modal discriminative representations while increasing inter-class distance and suppressing intra-class variation. Furthermore, to resolve the cross-modal alignment problem, we propose the cross-view instance projection matching to establish the relationship between different modal representations. Additionally, regarding how to effectively handle incomplete samples, we introduce nearest neighbor consistent completion, which can restore high-quality completion features. We adopt CLIP-ViT and CLIP-Xformer as image and text encoders respectively, and Rank-k metrics (k=1,5,10) as primary evaluation metrics. Extensive experiments on several public datasets demonstrate that our method can effectively address TIReID in open environments, consistently surpassing several SOTA TIReID methods.

Memory-aware continual learning with multi-modal social media streams for unsupervised disaster classification

Social media has emerged as a major hub for information dissemination during disasters. It serves as a central repository for massive real-time updates about disasters, facilitating the organization of humanitarian assistance. Disaster types classification, which identifies and categorizes various types of disasters on social media, has become a fundamental technique for rapid disaster response. However, current disaster classification methods suffer from the following two shortcomings: (1) They rely on manual annotation to train the models, which is expensive and time-consuming in rapidly updating social media; (2) During the disaster, reports on social media are progressively enriched from initial text to multi-modal mix containing image, audio, and video. Existing methods cannot handle these incremental multi-modal data. To address these challenges, we propose a Memory-aware Continual Learning (MaCL) method, which accomplishes unsupervised disaster classification by flexibly handling the multi-modal streams on social media in a continual learning manner. First, a modal-shared contrastive encoder is constructed to mine the correlation information among the textual and visual modalities on social media. We theoretically demonstrate that mining correlation information is approximately equivalent to preserving predictive information in the original data. Then, we design a Correlation Memory Bank (CMB) to store the features of current modalities. As new modalities emerge, the correlation information between the new modality and its global neighbors in the CMB is maximized. This strategy enables the correlation information to guide the classification of the newly arrived modalities. Finally, the new correlation information generated by the CMB continuously updates and accumulates, leading to improve disaster classification performance across all modalities. Extensive experiments on seven real-world datasets validate the effectiveness of the proposed MaCL.

AdaFN-AG: Enhancing multimodal interaction with Adaptive Feature Normalization for multimodal sentiment analysis

In multimodal sentiment analysis, achieving effective fusion among text, acoustic, and visual modalities for enhanced sentiment prediction is a crucial research topic. Recent studies typically employ tensor-based or attention-based mechanisms for multimodal fusion. However, the former fails to achieve satisfactory prediction performance, and the latter complicates the computation of fusion between non-textual modalities. Therefore, this paper proposes the multimodal sentiment analysis model based on Adaptive Feature Normalization and Attention Gating mechanism (AdaFN-AG). Firstly, facing highly synchronized non-textual modalities, we design the Adaptive Feature Normalization (AdaFN) method, which focuses more on sentiment features interaction rather than timing features association. In AdaFN, acoustic and visual modality features achieve cross-modal interaction through normalization, inverse normalization, and mix-up operations, with weights utilized for adaptive strength regulation of the cross-modal interaction. Meanwhile, we design the Attention Gating mechanism that facilitates cross-modal interactions between textual and non-textual modalities through cross-attention and captures timing associations, while the gating module concurrently regulates the intensity of these interactions. Additionally, we employ self-attention to capture the intrinsic correlations within single-modal features. Subsequently, we conduct experiments on three benchmark datasets for multimodal sentiment analysis, with the results indicating that AdaFN-AG outperforms the baselines across the majority of evaluation metrics. Through research and experiments, we validate that AdaFN-AG not only enhances performance by adopting appropriate methods for different types of cross-modal interactions while conserving computational resources but also verifies the generalization capability of the AdaFN method.

Cross-modal retrieval based on multi-dimensional feature fusion hashing

Along with the continuous breakthrough and popularization of information network technology, multi-modal data, including texts, images, videos, and audio, is growing rapidly. We can retrieve different modal data to meet our needs, so cross-modal retrieval has important theoretical significance and application value. In addition, because the data of different modalities can be mutually retrieved by mapping them to a unified Hamming space, hash codes have been extensively used in the cross-modal retrieval field. However, existing cross-modal hashing models generate hash codes based on single-dimension data features, ignoring the semantic correlation between data features in different dimensions. Therefore, an innovative cross-modal retrieval method using Multi-Dimensional Feature Fusion Hashing (MDFFH) is proposed. To better get the image’s multi-dimensional semantic features, a convolutional neural network, and Vision Transformer are combined to construct an image multi-dimensional fusion module. Similarly, we apply the multi-dimensional text fusion module to the text modality to obtain the text’s multi-dimensional semantic features. These two modules can effectively integrate the semantic features of data in different dimensions through feature fusion, making the generated hash code more representative and semantic. Extensive experiments and corresponding analysis results on two datasets indicate that MDFFH’s performance outdoes other baseline models.

Multimodal deep learning for dementia classification using text and audio

Dementia is a progressive neurological disorder that affects the daily lives of older adults, impacting their verbal communication and cognitive function. Early diagnosis is important to enhance the lifespan and quality of life for affected individuals. Despite its importance, diagnosing dementia is a complex process. Automated machine learning solutions involving multiple types of data have the potential to improve the process of automated dementia screening. In this study, we build deep learning models to classify dementia cases from controls using the Pitt Cookie Theft dataset from DementiaBank, a database of short participant responses to the structured task of describing a picture of a cookie theft. We fine-tune Wav2vec and Word2vec baseline models to make binary predictions of dementia from audio recordings and text transcripts, respectively. We conduct experiments with four versions of the dataset: (1) the original data, (2) the data with short sentences removed, (3) text-based augmentation of the original data, and (4) text-based augmentation of the data with short sentences removed. Our results indicate that synonym-based text data augmentation generally enhances the performance of models that incorporate the text modality. Without data augmentation, models using the text modality achieve around 60% accuracy and 70% AUROC scores, and with data augmentation, the models achieve around 80% accuracy and 90% AUROC scores. We do not observe significant improvements in performance with the addition of audio or timestamp information into the model. We include a qualitative error analysis of the sentences that are misclassified under each study condition. This study provides preliminary insights into the effects of both text-based data augmentation and multimodal deep learning for automated dementia classification.

A global multimodal flood event dataset with heterogeneous text and multi-source remote sensing images

ABSTRACT With the increasing frequency of floods, in-depth flood event analyses are essential for effective disaster relief and prevention. Satellite-based flood event datasets have become the primary data source for flood event analyses instead of limited disaster maps due to their enhanced availability. Nevertheless, despite the vast amount of available remote sensing images, existing flood event datasets continue to pose significant challenges in flood event analyses due to the uneven geographical distribution of data, the scarcity of time series data, and the limited availability of flood-related semantic information. There has been a surge in acceptance of deep learning models for flood event analyses, but some existing flood datasets do not align well with model training, and distinguishing flooded areas has proven difficult with limited data modalities and semantic information. Moreover, efficient retrieval and pre-screening of flood-related imagery from vast satellite data impose notable obstacles, particularly within large-scale analyses. To address these issues, we propose a Multimodal Flood Event Dataset (MFED) for deep-learning-based flood event analyses and data retrieval. It consists of 18 years of multi-source remote sensing imagery and heterogeneous textual information covering flood-prone areas worldwide. Incorporating optical and radar imagery can exploit the correlation and complementarity between distinct image modalities to capture the pixel features in flood imagery. It is worth noting that text modality data, including auxiliary hydrological information extracted from the Global Flood Database and text information refined from online news records, can also offer a semantic supplement to the images for flood event retrieval and analysis. To verify the applicability of the MFED in deep learning models, we carried out experiments with different models using a single modality and different combinations of modalities, which fully verified the effectiveness of the dataset. Furthermore, we also verify the efficiency of the MFED in comparative experiments with existing multimodal datasets and diverse neural network structures.