https://orcid.org/0000-0001-5103-2844
Figures
Abstract
Floods are among the devastating types of disasters in terms of human life, social and financial losses. Authoritative data from flood gauges are scarce in arid regions because of the specific type of dry climate that dysfunctions these measuring devices. Hence, social media data could be a useful tool in this case, where a wealth of information is available online. This study investigates the reliability of flood related data quality collected from social media, particularly for an arid region where the usage of flow gauges is limited. The data (text, images and videos) of social media, related to a flood event, was analyzed using the Machine Learning approach. For this reason, digital data (758 images and 1413 video frames) was converted into numeric values through ResNet50 model using the VGG-16 architecture. Numeric data of images, videos and text was further classified using different Machine Learning algorithms. Receiver operating characteristics (ROC) curve and area under curve (AUC) methods were used to evaluate and compare the performance of the developed machine learning algorithms. This novel approach of studying the quality of social media data could be a reliable alternative in the absence of real-time flow gauges data. A flash flood that occurred in the United Arab Emirates (UAE) from March 7–11, 2016 was selected as the focus of this study. Random forest showed the highest accuracy of 80.18% among the five other classifiers for images and videos. Precipitation/rainfall data were used to validate social media data, which showed a significant relationship between rainfall and the number of posts. The validity of the machine learning models was assessed using the area under the curve, precision-recall curve, root mean square error, and kappa statistics to confirm the validity and accuracy of the model. The data quality of YouTube videos was found to have the highest accuracy followed by Facebook, Flickr, Twitter, and Instagram. These results showed that social media data could be used when gauge data is unavailable.
Citation: Khan Q, Kalbus E, Zaki N, Mohamed MM (2022) Utilization of social media in floods assessment using data mining techniques. PLoS ONE 17(4): e0267079. https://doi.org/10.1371/journal.pone.0267079
Editor: Guy J-P. Schumann, Bristol University/Remote Sensing Solutions Inc., UNITED STATES
Received: November 4, 2020; Accepted: April 3, 2022; Published: April 25, 2022
Copyright: © 2022 Khan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data underlying the results presented in the study are available from the social media platforms (Twitter, Facebook, Instagram, Flickr and YouTube). We have included the Python code that we developed to collect/analyze this data in Appendix A.
Funding: This work was supported by the National Water and Energy Center at UAEU under funds # 31N319 and 31R191. The funder provided support in the form of salaries for authors QK and EK but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section.
Competing interests: The authors have declared that no competing interests exist.
Introduction
The arid climate of the Arabian Peninsula presents unique challenges for flood management [1]. Floods occurring after high-intensity rainfall events are a significant concern as flooding has considerable impacts including damage to infrastructure and loss of life [2]. Attempts to mitigate the impact of such incidents are, therefore, crucial especially with the predictions of increased climatic extremes associated with global climate change.
Rapid urbanization has altered the hydrological characteristics of land use, especially in the United Arab Emirates (UAE) [3]. The accumulation of surface water as a result of short-term rainfall is observed in the eastern and northern parts of the UAE [3, 4]. Flash floods in the UAE generally occur between November and March, where the mean runoff is about 120 Mm3 per year [5]. In the UAE, flow monitoring started in 1975 with seven flow gauges, which was increased to 21 in 1998 but then reduced to 10 flow gauges in 2005, and remained unchanged since then. For comparison, the City of Auckland in New Zealand operates 49 flow gauges over an area roughly the same size of Dubai [6].
To overcome the problem of flow data scarcity, the volunteered geographic information (VGI) approach provides the opportunity to collect additional information at low cost. The idea is to involve the public in the collection of relevant data according to pre-set standards and guidelines. This is also known as user-generated content (UGI). The use of VGI in flood-related studies started recently [7], with researchers estimating flood damage from VGI [8] or mapping flood extents [9, 10].
Recent approaches take advantage of the wealth of data distributed through social networks. Social media messages that have a geographic reference can be considered VGI, as they can be used to analyze what is happening at a specific location [11]. Participation by citizens is implicitly volunteered as they do not distribute information for flood monitoring [12]. Various social media platforms have been used to extract and analyze flood-related information including Twitter [13–15], Facebook [15], Flickr [16], and YouTube [17]. Social media such as Twitter has also been used to study food-security related data during natural hazards [18].
The main advantages of these new data sources are (i) they can provide much denser geographical coverage compared to traditional sensor networks, and (ii) they directly record the impact of a flood on the human environment, as the users usually document personal observations and experiences [16]. For these reasons, majority of studies on the use of social media for flood monitoring have been conducted in the context of disaster response and flood damage assessment [13–15, 19–22].
Due to the absence of flow gauges and appropriate quality data in arid regions, the identification of other sources of data is essential. This study evaluates the use of social media data related to floods, which has never been studied before in the GCC region. The quality of social media data, in addition to the data of each individual social media platform, Twitter, Facebook, Instagram, YouTube, and Flickr, was investigated.
Literature review
Using social media-derived flood information has great potential for hydrological research to improve flood management. Photos with geographical reference and time stamps taken at different times during a flood event can be used to estimate the development of the flood hydrograph and the flow rate (in conjunction with other technical data from the catchment). Videos can show the onset of a flash flood and the movement of the flood wave through the catchment. Such data can be valuable for calibrating flow models. However, only a few studies have examined the use of this type of data for hydrological research and modelling [23–25]. Even fewer studies have investigated the use of social media data for the development of monitoring records for ungauged water bodies [17]. Examples of such studies are presented below:
Barker and Macleod [26] studied Twitter data to monitor flood events using real-time river levels across the United Kingdom. A paragraph vectors and a logistic regression-based classifier was used in the development of a Twitter data mining pipeline. The flood data obtained were then integrated with a real-time environmental data to give stakeholders better understanding of the local conditions. Another study by Bischke et al. [27] used satellite images for flood detection and enriched the information by using images from Twitter to understand the severity of floods. Twitter images were pre-processed by removing duplicate images and then an algorithm was designed that can identify those images and the water level as a result of flood, secondly high-resolution satellite images were also studied for high flood level by automatically detection of water levels and generating flood map. To identify the relevancy of georeferenced social media messages from Twitter during any flood event, de Albuquerque et al. [11] used statistical analysis, for identification of spatial patterns in the flood-related tweets and combining it with authoritative data by analyzing a case study of the River Elbe Flood in Germany in June 2013. The results showed that Tweets within a spatial distance of 10 km has higher probability of being related to a specific flood incident. The approach of the “wisdom of the crowd” with the number of Tweets—reliable patterns in the data, for a certain event versus the monitored watershed activity was studied for Jakarta, Indonesia [19]. This approach is more applicable to areas where more social media users are active. The observed information about the georeferenced activity of Twitter users in an area was mapped on a Digital Elevation Model (DEM) with flood depth observations and then used to create flood maps. The tweets with water depth showed very good and reliable indication of flood in that area. Similar study, by mapping flood depth from social media data, was also conducted by Karmegam et al. [28], and the results were validated with real time flow data.
A more detailed study in Argentina, France, and New Zealand was conducted by Le Coz et al. [23], where a dedicated website of the Flood Chasers Project was developed. People were encouraged to upload images and videos of any flood events. These images were processed using PIV/PTV analysis tools for flood mapping, and flood estimation using videos of river flow velocity and discharge were modeled using Large Scale Particle Image Velocimetry (LSPIV), which is an effective technique for post-flood discharge estimation. The implementation of these tools provides potential of citizen science for flood risk assessment. Other such studies in literature has been reported that uses VGI focused on flood damage using dedicated platforms, such as PetaJakarta in Jakarta, QLF flood crisis map in Australia [29], and flooding points in Brazil [30].
Restrepo-Estrada et al. [31] studied the social media for rainfall-runoff estimations and forecasting floods using data from Twitter and authoritative data. They combined geospatial Twitter data with real-time authoritative flood values as an input for the Probability Distribution Model (PDM) and achieved 71% accuracy. Addition of social media data has increased the accuracy of PDM almost twice. Rosser et al. [22] studied the 2014 UK flood using Flickr posting activity, remote sensing, and topographic map data using a Bayesian statistical model to develop a probability map that shows the likelihood of the presence of floodwater.
Panigrahi et al. [32] used deep neural networks for prediction of floods in terms of seven parameters including rainfall, area, pressure, velocity, gauge, average temperature and average windspeed. Two rivers Daya and Bhargavi in India were studied using deep learning models for predicting discharge volume. Results indicated that Local Linear Radial Basis Functional Neural Network (LLBRFNN) performed well in prediction of flood levels with lowest Mean Square error (MSE) and Mean Absolute Percentage Error (MAPE). Same rivers were used in another study by Panigrahi et al. [33] where Cascaded Functional Link Artificial Neural Network (C-FLANN) was developed; Harmony Search (HS) and Differential Evolution (DE) was used to update the parameters of the model. Same seven atmospheric parameters were used to predict water flow in the rivers where the results showed that c-FLANN trained using HS gives better predictions of water level.
The major task in using social media for flood monitoring is the classification of large amounts of data in various formats (image, video, and text) and the extraction of relevant parameters from the data. Typically, machine-learning methods are used to aid in this task. For example, Huang et al. [34] used unsupervised machine learning method (clustering) to cluster text related to emergency based on similarity and combine it with supervised logistical regression to cluster posts based on different events. The selected methods for this study are described below.
ResNet is a pre-trained model for image classification [35]. The ResNet architecture is often used for feature extraction, image classification, image segmentation, and object detection. This method uses deep convolutional neural networks (CNNs) for identification and classification purposes.
Random forest is a machine-learning algorithm that is used for classification and regression. It is a method that constructs multiple decision trees. During the training time, individual decision trees are generated by randomly selecting the attributes in each node that determines the split. Each tree then gives its separate weights individually during classification. Random forests can handle data with outliers, and it can also handle missing values [36].
Naïve Bayes is a statistical classifier that provides conditional independence between predictors [36]. It is mainly based on the assumption that all predictors or attributes are conditionally independent, which is the reason for its naivety [37].
Sequential minimal optimization (SMO) is an implementation of a support vector machine (SVM) classifier in the Weka (Waikato Environment for Knowledge Analysis) platform. It is developed for numeric prediction and classifying data by the construction of an N-dimensional hyperplane that can separate data optimally into two categories [38]. SVM works well in text classification tasks as it has the ability to remove the need for feature selection [39].
Classification algorithm C4.5 produces a decision tree based on information theory. It uses the information entropy concept and uses a greedy technique to induce decision trees for classification [40]. It accepts nominal classes and is used in the construction of a decision tree from labelled training data that uses information entropy [37].
Earlier studies have focused on the development of a framework for social media activity in flooded areas, and some of them used dedicated volunteer geographic information to study such events on the hydrological catchment scale. However, there is a need to address the data quality of collected information from social media and the validity of such data for flood predictions, particularly for arid regions with patchy but intense rainfall and high risk of flash floods. This study, for the first time, analyses data obtained in an arid region from various social media platforms such as Twitter, Facebook, Instagram, Flickr, and YouTube. It validated social media data for flood monitoring. For this purpose, social media data from a period of heavy rainfall and the resulting flash flood events (March 7 to 11, 2016) were analyzed for the UAE, which is classified as an arid region. The validation of social media data for flood prediction and evaluation was performed based on the outputs of different machine learning classifiers. For this purpose, the area under the curve (AUC), root-mean-square error (RMSE), and kappa statistics were used. Data quality from Twitter, Facebook, Instagram, YouTube, and Flickr were also analyzed, and their model performance was assessed using AUC, RMSE, and kappa statistics.
Data collection
The UAE is located in a tropical dry region, where the Tropic of Cancer runs to the south of the country. The climate in the UAE is characterized by high temperatures throughout the year, [41]. Precipitation is scarce, and rainfall occurs only from December to April [42]. The average annual rainfall in 2015 was 87.4 mm, which decreased to 60.7 mm in 2016 (Fig 1).
Mild rainfall (1.2 mm) in 2016 occurred on March 8, which was followed by heavy rainfall (7.3 mm) on March 9 that brought the life to standstill in the UAE. To study the flash flood that occurred due to heavy rainfall in the UAE, 794 online posts were collected from different platforms using keywords and geographical queries given in Table 1. These hashtags were used in combination with the operators (‘AND’ and ‘OR’). To obtain relevant data on the UAE and seven different emirates, each geographical query (UAE, Abu Dhabi, Dubai, Al Ain, Sharjah, and RAK) was used with ‘AND’ followed by rain, flood, storm, and weather with ‘OR’ operator. Three different types of data were collected: text, images, and videos. To validate the data collected from social media, precipitation/rainfall data (mm/day) was obtained from the Global Precipitation Measurement (GPM), which is a climatic satellite that has the ability to detect and measure precipitation using advanced instruments. GPM data for the time frequency of day were obtained from the National Aeronautics and Space Administration (NASA) website (https://pmm.nasa.gov/data-access/downloads/gpm) for the period of March 7 to 11, 2016. A study by Mahmoud et al. [43] validated this data for the UAE for the period from 2015 to 2017 using ground values from the National Centre of Meteorology (NCM).
The text, images and videos obtained from studied social media platform were then used to access the flash flood using machine-learning algorithms. The parameters used for carrying out simulations contain three major categories of text, images and videos. The text messages were converted to binary matrix which was then used as an input into the machine learning simulations. However, videos were first converted into frames, using python code (S1 Appendix), and those frames along with images were converted into features using VGG-16 architecture and then used as input data into machine learning model. Additionally, for simulation, 20-fold cross validation was used.
Methodology
The pre-processing of data includes categorization of data into four classes: not relevant, rain, low flood, and high flood (Table 2). All the texts, images, and videos were categorized based on these classes. Duplicate images and videos were removed based on the user who first uploaded them on social media.
Fig 2 shows the sample images from the complete dataset, which were manually classified based on four classes (Table 2). The “irrelevant” class was assigned to images/videos that were irrelevant to our study but used the hashtags given in Table 1. For example, users who uploaded images of landmark buildings, clouds, swimming pools, and advertisements (commercial companies use weather-trending hashtags to sell food products during good weather) were assigned an irrelevant class.
Sample images from final dataset containing classified images based on four classes (a) not relevant, (b) rain, (c) low flood and (d) high flood.
The ResNet50 model is pre-trained on the ImageNet dataset and is available with the Keras API. ResNet50 was used as a transfer learning technique for CNNs to extract features. A Python code (S1 Appendix) was developed to (1) convert videos into frames and (2) extract features from converted frames and images.
Feature extraction is based on VGG and ResNet architectures [44]. The VGG-16 feature extractor was initialized with the pertained weights created using ImageNet [45]. The architecture of VGG-16 has been largely used in visual data classification consisting of stacked convolution and max pooling layers (Fig 3). The input to the first layer is an RGB image with a size of 224 × 224. The image is then passed through different layers, which have filters with a very small receptive field of 3 × 3. The final layer is the SoftMax layer, which contains 1000 channels.
This process of feature collection for the identification of rain and floods in the images was executed using the programming of Keras with TensorFlow at the backend in the Anaconda Python environment. This was carried out to obtain features that could distinguish between different classes of our study. Each image and frame in the case of a video were converted into 1000 flattened features.
The text messages associated with the videos and frames were also converted into a binary coded matrix for analysis in Weka. As the text message was uploaded on social media with image/video, they were organized in the same row along with their class (first column), which was assigned manually during image and video categorization (Fig 4).
The features obtained from images and frames of video along with the text messages were classified based on four classes using the Weka tool (Fig 5). The data mining tool Weka, which was developed in the Java language was used in this study. This tool has many data-mining algorithms that are grouped into different groupings according to the rules generated by the algorithm. For this study, we used classification algorithms, namely random forest, k-nearest neighbors (IBk), naïve Bayes, support vector machine (SMO), and C4. 5 (J48). The test mode used was 20-fold cross-validation, and “full training set” was used as classifier model. The classification method involved 20-fold repetitions of the validation process 20 times.
The initial validation of study was conducted by comparing the activity of social media users with the precipitation/rainfall data from Global Precipitation Measurement (GPM). GPM, a climatic satellite, has many products, which are categorized into four levels by NASA. Level 0, 1, 2 and 3. Level-0 is raw unprocessed data which is used by Level-1 and produces brightness temperature. Level-2 uses Level-1 data and produces rainfall estimates. Whereas, Level-3 is recommended for use, as it is provided by Integrated Multi-satellite Retrievals (IMERG) algorithms and it combines all microwave, infrared satellite estimates and ground precipitation gauges [43]. Hence, this study used IMERG products (Level-3) data of precipitation. Since the data provided by IMERG is half-hourly precipitation (mm/30 min), the IMERG products were downloaded by accumulating half-hourly data to daily estimates for the period of March 7 to 11, 2016.
Results and discussion
A total of 2171 rows containing 758 images and 1413 frames extracted from 154 videos, each having 1000 columns of attributes were analyzed and categorized into classes. Additionally, 705 texts from different online posts associated with these videos and images were also analyzed after converting them into a binary matrix. Different classifiers were used to test the data based on four different classes (irrelevant, rain, low flood, and high flood).
The model was set to run for three instances, first on the attributes of images and videos, then on the attributes of images and videos along with text, and finally, only on text. Table 3 shows the model accuracy and the time taken to build the model.
Result showed an interesting pattern where random forest gained the highest accuracy (80.18%), and Naïve Bayes had the lowest accuracy (37.83%) for the attributes of images and videos (Table 1). The text showed an accuracy of 61.28% using random forest, and the highest accuracy of 63.12% using C4. 5 (J48) classifier. When the text messages were combined with the attributes of video images, the accuracy of random forest dropped to 64.3%. This shows that the comparatively lower classified instances of text have reduced the accuracy from 80.18% to 64.3%.
For the execution time, k-nearest neighbors required 4E-3, 2E-2 and 1E-3 seconds for ‘images & videos’, ‘text, images & videos’ and ‘text’ respectively. The highest time taken to build the model was the C4. 5 (J48) as 11.69 s for images and videos whereas for the random forest, which showed highest accuracy for images and videos, took 3.78 s to build the model.
On the basis of accuracy and time while studying three instances for flood prediction, it was confirmed that random forest is the best classifier that gained the highest accuracy and comparatively less time to build the model on image and video attributes. This also showed that the information provided by users on social media in terms of text messages is less relevant. Additionally, the Naïve Bayes algorithm showed the least accuracy among all classifiers for all three instances, and hence, it cannot be used for the purpose of such classifications.
The scarcity of flow data is the result of a combination of conditions typically observed in arid areas, especially in the UAE. Almost all water courses on the Arabian Peninsula are ephemeral or intermittent in nature with water flow only for short periods of time. The region is generally sparsely populated (except in coastal areas), which makes regular maintenance visits as well as event-based visits challenging. Flood events typically occur in the form of flash floods because of high-intensity, short-duration rainfall events. Furthermore, the flood volume is often very high, which may cause the destruction of the flow gauge blocking of the gauge with large debris carried with the flood, or simply causing conditions that are too dangerous for manual flow gauges. Climatic conditions with extremely high temperatures during the summer may cause failure of the monitoring equipment.
The rainfall data obtained from GPM for the UAE showed that rainfall started from March 7, 2016 was 0.63 mm and was the highest on March 9, 2016 at 7.3 mm. For the validation of data obtained from social media, two time series were plotted against the date. The number of online posts collected for this study from social media platforms were plotted on time series with the total number of posts (including text, images, and videos). Only ‘images and videos’ uploaded on selected dates are plotted against intensity of rainfall in mm. Evidence obtained from plotting these time series showed a significant relationship between rainfall and the frequency of posts, as well as the uploaded images and videos (Fig 6).
Time series of rainfall depths (a) with frequency of total posts per day, (b) with frequency of images and videos per day.
The actual rainfall event that triggered the flash flood was on March 9 and the frequency of posts was highest at 485. The same pattern was also observed for the images and videos uploaded on social media platforms, which were 351. This showed that social media users were more active on the day of the highest precipitation event. The activity on the hashtags of rain, flood, storm, and weather was highest on social media for this particular geographic location and the event.
Random Forest achieved the highest accuracy of 80.18% for the attributes of images and videos. Nair et al. [36] studied the frequency of social media users on the day of a flood in India and achieved the highest accuracy of 99.9% using random forest. However, the quality of social media data is never used before for studying floods. This study showed with reasonable classification accuracy, especially through Random Forest classifier, that machine learning algorithms can be effectively used to access visual and textual social media data for flood study.
Models validation
The random forest performed best among all five classifiers, and the evaluation of all models is presented in Table 4. The kappa statistics showed the reliability of the model and that classification does not occur by chance [46]. The value of kappa statistics is 0.63, which is highest for images and video classification of random forest followed by KNN.
The root mean square error (RMSE) shows the difference between the observed values and model-predicted values [47]. The highest RMSE value (0.56) was recorded by Naïve Bayes and the least by Random Forest in image and video classification. Model performance can be evaluated using the AUC, with values lower than 0.5 indicating the inefficiency of the model, and values greater than 0.8 are considered good models [48]. The data formats of ‘images & videos’ and ‘images, videos, and text’ for the random forest classification showed the highest 0.8 AUC values, also plotted in Fig 7.
Since model evaluation only through AUC can be deceiving, precision recall (PRC) values are used to evaluate the sensitivity of the model along with positive predictive values [47]. The PRC value for ‘images and videos’ of random forest was the highest. These results showed that the accuracy and validity of random forest for images and videos were the highest among all other classifiers and data formats.
Data quality assessment
The quality of the data obtained from social media is questionable. For instance, social media users usually add more hashtags, which could contain rain and floods in the same text, to get more viewership. This is reflected in the accuracy of this study, where text messages have comparatively low accuracy compared to images and videos. Additionally, other data quality issues are related to ‘irrelevant’ images and videos shared. For example, sharing an image of clouds or swimming pools with the hashtags of rain and floods reveals discrepancies. Also, social media users are not so technical to distinguish between low flood or high flood, which is replicated in the accuracy being comparatively lower than ‘images & videos’. To understand the quality of data on different social media platforms, data from Twitter, Facebook, Instagram, YouTube, and Flickr were evaluated separately using the random forest classifier (Table 5).
Video data from YouTube classified best with 83.61% accuracy followed by 80.46% for ‘image & videos’ data from Facebook. The AUC and PRC values for YouTube videos were also highest at 0.96 and 0.93, respectively. However, the classification of ‘Text’ from YouTube (video caption and description) achieved the lowest accuracy of 43.18% among all formats.
This is the first study in an arid region that analyzed data quality along with social media behavior usage during such flood events. The detailed social media data quality analysis—in terms of flood study, of each studied social media platform, under different formats of text, images and videos is explained in detail. Currently, there are no active flow monitors in the UAE. However, such flash floods are more common and are triggered by heavy precipitation. Hence, in such cases, the images and videos obtained from social media can be relied upon. Considering that flood data are very scarce across the UAE, such a database will be extremely valuable for public authorities concerned with flood management.
Conclusions
An alternative method for flood analysis is suggested in this research, by proposing the use of social media data. This study aimed to investigate the quality of social media data for flood monitoring. Data related to flood events in the UAE of March 7 to 9, 2016 were collected from Facebook, Twitter, Instagram, YouTube, and Flickr. Results showed that Random Forest achieved highest accuracy with 80.18% for the data format of ‘images and videos.’ The binary codes of the text showed the least accuracy of 61.28%. The data from social media showed a significant correlation with rainfall data from the general circulation model and the number of posts mentioning flood-related keywords. The performance of the machine learning models was evaluated using the AUC and PRC. Random Forest also showed the highest AUC of 94% and PRC of 88% for ‘images and videos’. The data quality showed that Random Forest classified YouTube data of videos with highest accuracy, followed by ‘images & videos’ of Facebook, Flickr, Twitter and Instagram. These results indicated that the quality of images and videos from social media can be used for flood monitoring purposes by public authorities, especially in areas of the UAE with no active flow gauges.
The collection stage of the data was extremely intensive and time consuming due to the use of multiple hashtags and on different social media platforms. The application developed in S1 Appendix can be further enhanced by automating the input of data from social media platforms using different Application Programming Interface (APIs). Furthermore, the study of quality of social media data can be further advanced by using artificial intelligence and deep learning models to measure the flood water level in images and videos.
Supporting information
S1 Appendix. Python code for frames extraction from videos and conversion of images into features.
https://doi.org/10.1371/journal.pone.0267079.s001
(DOCX)
Acknowledgments
Special thanks go to Serin Darwish and Srikanth Bezawada from UAEU for their help in the pre-processing stage of data.
References
- 1. Kotwicki V, Al Sulaimani Z. Climates of the Arabian Peninsula–past, present, future. Int J Clim Change Strateg Manag. 2009;1: 297–310. https://doi.org/10.1108/17568690910977500.
- 2. Jonkman SN. Global perspectives on loss of human life caused by floods. Nat Hazards. 2005;34: 151–175. https://doi.org/10.1007/s11069-004-8891-3.
- 3. Shanableh A, Al-Ruzouq R, Yilmaz AG, Siddique M, Merabtene T, Imteaz MA. Effects of land cover change on urban floods and rainwater harvesting: a case study in Sharjah, UAE. Water. 2018;10: 631. https://doi.org/10.3390/w10050631.
- 4. Ahmed AA. An overview of conventional and non-conventional water resources in arid region: assessment and constrains of the United Arab Emirates (UAE). J Water Resour Prot. Scientific Research Publishing. 2010; Volume 02: 181–190. https://doi.org/10.4236/jwarp.2010.22020.
- 5. Sherif MM, Mohamed MM, Shetty A, Almulla M. Rainfall-runoff modeling of three wadis in the northern area of UAE. J Hydrol Eng. American Society of Civil Engineers. 2011;16: 10–20. Available form. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000280.
- 6.
AC. The Health of Auckland’s Natural Environment in 2015 [Internet]. Auckland Council; 2015. Available: https://www.aucklandcouncil.govt.nz/environment/state-of-auckland-research-report-cards/Pages/health-of-aucklands-natural-environment-2015.aspx.
- 7. Goodchild MF. Citizens as sensors: the world of volunteered geography. GeoJournal. 2007;69: 211–221. https://doi.org/10.1007/s10708-007-9111-y.
- 8.
Poser K, Kreibich H, Dransch D. Assessing Volunteered Geographic Information for Rapid Flood Damage Estimation; 2009. Available: https://gfzpublic.gfz-potsdam.de/pubman/item/item_239400. Leibnitz University Hannover, Germany.
- 9. McDougall K, Temple-Watts P. The use of LiDAR and volunteered geographic information to map flood extents and inundation. ISPRS Ann Photogramm Remote Sens Spatial Inf Sci. 2012; Volume I–4: 251–256. https://doi.org/10.5194/isprsannals-I-4-251-2012.
- 10. Triglav-Čekada M, Radovan D. Using volunteered geographical information to map the November 2012 floods in Slovenia. Nat Hazards Earth Syst Sci. 2013;13: 2753–2762. https://doi.org/10.5194/nhess-13-2753-2013.
- 11. Albuquerque JP de, Herfort B, Brenning A, Zipf A. A geographic approach for combining social media and authoritative data towards identifying useful information for disaster management. Int J Geogr Inf Sci. 2015;29: 667–689. https://doi.org/10.1080/13658816.2014.996567.
- 12. Assumpção TH, Popescu I, Jonoski A, Solomatine DP. Citizen observations contributing to flood modelling: opportunities and challenges. Hydrol Earth Syst Sci Discuss. 2017:22:1473–1489. https://doi.org/10.5194/hess-22-1473-2018.
- 13.
Ashktorab Z, Brown C, Nandi M, Culotta A. Tweedr: Mining twitter to inform disaster response; 2014. Proc 11th Int ISCRAM Conf [Internet]. Pennsylvania, USA. Available: http://cs.iit.edu/~culotta/pubs/ashktorab14tweedr.pdf. [Accessed 2017 Sep 26].
- 14. Cheong F, Cheong C. Social Media Data Mining: A Social Network Analysis Of Tweets During The 2010–2011 Australian Floods. PACIS. 2011 Proc;2013. Paper p. Available: https://aisel.aisnet.org/pacis2011/46/46.
- 15. Spielhofer T, Greenlaw R, Markham D, Hahne A. Data mining Twitter during the UK floods: Investigating the potential use of social media in emergency management 3rd Int Conf Inf Commun Technol Disaster Manag ICT-DM. 2016; 2016. pp. 1–6. https://doi.org/10.1109/ICT-DM.2016.7857213.
- 16. Tkachenko N, Jarvis S, Procter R. Predicting floods with Flickr tags. PLOS ONE. 2017;12: e0172870. pmid:28235035
- 17. Michelsen N, Dirks H, Schulz S, Kempe S, Al-Saud M, Schüth C. YouTube as a crowd-generated water level archive. Sci Total Environ. 2016;568: 189–195. pmid:27295591
- 18. Niles MT, Emery BF, Reagan AJ, Dodds PS, Danforth CM. Social media usage patterns during natural hazards. PloS One. Public Library of Science San Francisco, CA USA; 2019;14:e0210484. pmid:30759111
- 19. Eilander D, Trambauer P, Wagemaker J, van Loenen A. Harvesting social media for generation of near real-time flood maps. Procedia Eng. 2016;154: 176–183. https://doi.org/10.1016/j.proeng.2016.07.441.
- 20. Fohringer J, Dransch D, Kreibich H, Schröter K. Social media as an information source for rapid flood inundation mapping. Nat Hazards Earth Syst Sci. 2015;15: 2725–2738. https://doi.org/10.5194/nhess-15-2725-2015.
- 21. Kalas M, Kliment T, Salamon P. Coupling flood forecasting and social media crowdsourcing; 2016 [Cited 2017 Sep 26]. Available: http://adsabs.harvard.edu/abs/2016EGUGA.1815865K. p. 15865.
- 22. Rosser JF, Leibovici DG, Jackson MJ. Rapid flood inundation mapping using social media, remote sensing and topographic data. Nat Hazards. 2017;87: 103–120. https://doi.org/10.1007/s11069-017-2755-0.
- 23. Le Coz J, Patalano A, Collins D, Guillén NF, García CM, Smart GM, et al. Crowdsourced data for flood hydrology: Feedback from recent citizen science projects in Argentina, France and New Zealand. J Hydrol. 2016;541: 766–777. https://doi.org/10.1016/j.jhydrol.2016.07.036.
- 24. Smith L, Liang Q, James P, Lin W. Assessing the utility of social media as a data source for flood risk management using a real-time modelling framework. J Flood Risk Manage. 2017;10: 370–380. https://doi.org/10.1111/jfr3.12154.
- 25. Starkey E, Parkin G, Birkinshaw S, Large A, Quinn P, Gibson C. Demonstrating the value of community-based (‘citizen science’) observations for catchment modelling and characterisation. J Hydrol. 2017;548: 801–817. https://doi.org/10.1016/j.jhydrol.2017.03.019.
- 26. Barker JLP, Macleod CJA. Development of a national-scale real-time Twitter data mining pipeline for social geodata on the potential impacts of flooding on communities. Environ Modell Softw. 2019;115: 213–227. https://doi.org/10.1016/j.envsoft.2018.11.013.
- 27. Bischke B, Helber P, Zhao Z, De Bruijn J, Borth D. The multimedia satellite task at MediaEval 2018: Emergency response for flooding events. Work Notes. Proc Mediaev Workshop Mediaev; 2018. Ceur WS; 2018. org. Available: https://www.alexandria.unisg.ch/id/eprint/258247. pp. 1–3.
- 28. Karmegam D, Ramamoorthy S, Mappillairaju B. Near real time flood inundation mapping using social media data as an information source: a case study of 2015 Chennai flood. Geoenvironmental Disasters. Springer; 2021;8:1–11. https://doi.org/10.1186/s40677-021-00195-x.
- 29. Koswatte S, McDougall K, Liu X. XSDI and crowdsourced spatial information management automation for disaster management. Surv Rev. 2015;47: 307–315. https://doi.org/10.1179/1752270615Y.0000000008.
- 30. Hirata E, Giannotti MA. Larocca APC, Quintanilha JA. Flooding and inundation collaborative mapping–use of the Crowdmap/Ushahidi platform in the city of Sao Paulo, Brazil. J Flood Risk Manag. 2018:11: S98–S109. https://doi.org/10.1111/jfr3.12181.
- 31. Restrepo-Estrada C, de Andrade SC, Abe N, Fava MC, Mendiondo EM, de Albuquerque JP et al. Geo-social media as a proxy for hydrometeorological data for streamflow estimation and to improve flood monitoring. Comput Geosci. 2018;111: 148–158. https://doi.org/10.1016/j.cageo.2017.10.010.
- 32. Panigrahi BK, Nath TK, Senapati MR. An application of local linear radial basis function neural network for flood prediction. J Manag Anal. Taylor & Francis; 2019;6:67–87. https://doi.org/10.1080/23270012.2019.1566033.
- 33. Panigrahi BK, Das S, Nath TK, Senapati MR. An application of data mining techniques for flood forecasting: application in rivers Daya and Bhargavi, India. J Inst Eng India Ser B. Springer; 2018;99:331–42. https://doi.org/10.1007/s40031-018-0333-9.
- 34. Huang L, Shi P, Zhu H, Chen T. Early detection of emergency events from social media: A new text clustering approach. Nat Hazards. Springer; 2022;1–25. pmid:35095194
- 35. Gallego A-J, Pertusa A, Gil P. Automatic ship classification from optical aerial images with convolutional neural networks. Remote Sens. 2018;10: 511. https://doi.org/10.3390/rs10040511.
- 36. Nair MR, Ramya GR, Sivakumar PB. Usage and analysis of Twitter during 2015 Chennai flood towards disaster management. Procedia Comput Sci. 2017;115: 350–358. https://doi.org/10.1016/j.procs.2017.09.089.
- 37. Aharwal RP. Evaluation of various classification techniques of weka using different datasets. Int J Adv Res Innov Ideas Educ. 2016;2;4396: 2395.
- 38. Ayodele TO. Types of machine learning algorithms. New Adv. Mach Learn. IntechOpen. 2010. https://doi.org/10.5772/9385.
- 39.
Alshutayri A, Atwell ES, Alosaimy A, Dickins J, Ingleby M, Watson J. Arabic language WEKA-based dialect classifier for Arabic automatic speech recognition transcripts. Available: http://eprints.whiterose.ac.uk/107396/. Proc Third Workshop NLP Similar Lang Var Dialects VarDial 2016. Leeds; 2016. pp. 204–211.
- 40.
Quinlan JR. programs for machine learning. Available: https://ci.nii.ac.jp/naid/10015645285/. C4.5. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA; 1993.
- 41. Sheppard CRC. Physical environment of the Gulf relevant to marine pollution: an overview. Mar Pollut Bull. 1993;27: 3–8. https://doi.org/10.1016/0025-326X(93)90003-3.
- 42. Glennie KW, Singhvi AK. Event stratigraphy, paleoenvironment and chronology of SE Arabian deserts. Quat Sci Rev. 2002;21: 853–869. https://doi.org/10.1016/S0277-3791(01)00133-0.
- 43. Mahmoud MT, Hamouda MA, Mohamed MM. Spatiotemporal evaluation of the GPM satellite precipitation products over the United Arab Emirates. Atmos Res. Elsevier. 2019;219: 200–212. https://doi.org/10.1016/j.atmosres.2018.12.029.
- 44. Fuentes A, Yoon S, Kim SC, Park DS. A robust deep-learning-based detector for real-time tomato plant diseases and pests recognition. Sensors (Basel). 2017;17: 2022. pmid:28869539
- 45. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. Arxiv Prepr ArXiv. 2014. Available: https://arxiv.org/abs/1409.1556:14091556.
- 46. Monserud RA, Leemans R. Comparing global vegetation maps with the Kappa statistic. Ecol Modell. 1992;62: 275–293. https://doi.org/10.1016/0304-3800(92)90003-W.
- 47. Rizeei HM, Pradhan B, Saharkhiz MA, Lee S. Groundwater aquifer potential modeling using an ensemble multi-adoptive boosting logistic regression technique. J Hydrol. Elsevier. 2019;579. https://doi.org/10.1016/j.jhydrol.2019.124172, 124172.
- 48. Sajedi-Hosseini F, Malekian A, Choubin B, Rahmati O, Cipullo S, Coulon F, et al. A novel machine learning-based approach for the risk assessment of nitrate groundwater contamination. Sci Total Environ. Elsevier. 2018;644: 954–962. pmid:30743892