To validate the results obtained from the news articles, the events were compared to Water Storage Deficit (WSD) derived from the GRACE satellites. The indicator was used as a benchmark to the drought physical assessment. This choice was motivated by the opportunity to allow an integrated assessment of each component of the hydrological system. Ultimately, the results were moderately satisfying. The comparison with the WSD showed a high recall and low precision (Table 2). This suggests that although most droughts are identified, numerous false alarms are observed. Three challenges faced explains the low precision. First, we encountered the difficulty to properly match each event to their respective locations and dates within articles mentioning several dates and locations. Secondly, numerous articles showed coarse temporal expressions which would then result in lengthier drought events. Finally, we noticed an underestimation of the droughts durations from our benchmark drought indicator. 

Table 2: Skills of the news article-based detection compared to the Water Storage Deficit (WSD) in Mali

Immediate improvement and application areasThis research constitutes the first attempt to detect and monitor drought using online news articles. Although the results presented several shortcomings, future development remains promising. Among on-going improvement, we are aiming to use a machine learning framework to match each event their respective additional details, such dates,  locations and impacts. This upgrade will allow to improve the accuracy of the events reported. Potential application areas include:

• Disaster response organizations and water management organisations (Protection Civile, the Red Cross and Direction Nationale de l’Hydraulique) who can better respond to disasters such as droughts and water scarcity. These organizations lack actionable (near)-real-time data to request funds, shape their response and prioritize actions. 
• Agricultural production support organisations: Agricultural credit suppliers and insurance firms currently lack detailed, trustworthy and cost-efficient information regarding farmers' assets and related risks. Consequently, these financial actors cannot profitably enter the market of small-scale agricultural business. With the on-the-ground information on drought impact, uncertainties are reduced, and farmers would get better access to credit.
• Agro-meteorological information NGOs (such as MyAgro): The Malian population relies on small-scale, rainfed agriculture and pastoralism. Several applications have been developed to provide the farmers with detailed weather services. Timely information from online media on extreme weather impacts can complement these services and assist both farmers and pastoralists to improve food security. initiative aiming to develop impact forecast method in Mali. In this collaborative project, FloodTags provided the online media analysis and Satelligence made vegetation health index (NDVI) to study historic drought events. Deltares integrated both impact data to allow to create an impact forecast model. Finally, Akvo developed a reporting application for mobile devices to improve measurements communication from meteorological stations.

Case Study B: Twitter for drought detection and monitoring in California

In recent years, social media drew interest as a possible new source of information on natural disasters such as wildfire, earthquake, floods, winter storms and heavy snowfalls. However, it was never used yet for the detection and monitoring of droughts.

ObjectivesThe study investigated the use of Twitter to detect and monitor drought events and their impacts from a human–centric perspective. For this goal, it is first necessary to understand how tweets relates to the physical parameters of drought. The Water Storage Deficit derived from the GRACE satellites was chosen as a drought indicator, thus, as the reference to the drought physical assessment. This choice was motivated by the opportunity to allow an integrated assessment of each component of the hydrological cycle. Due to the elusiveness of the drought phenomena, and the diverse drivers and impacts it may encompass, a large set of preselected drought-related keywords were monitored and analysed to define which keyword is the most suitable to encompass the drought phenomena. Three keywords categories were reviewed (Table 3): drought identifier, drivers and impacts. 

Table 3: Preselected drought-related keywords

First, the tweets containing drought-related keywords were assigned a location using TAGGS, a geoparsing algorithms that enhances location disambiguation using both metadata and the contextual information of groups of tweets [3]. Then, the tweets were filtered to exclude all tweets that did not mention California or a geographic entity therein. Secondly, a blacklist was created to filter tweets containing the word “drought” in different contexts. Finally, the tweets monthly variations were compared to the GRACE indicator using an F-score.

ResultsThe F-score obtained are displayed in Table 4. The results validating the method was beyond expectation (82%), accounting for 44 months of drought detected by the online media system, where 39 months were confirmed by the referenced drought indicator (Water Storage Deficit). 

Table 4: Skills of the tweet-based detection compared to the Water Storage Deficit (WSD) in California (USA)

The visualisation of the tweets time series and the drought indicators can be found in Figure 2 and Figure 3. The tweets showed an exponential increase during GRACE-detected drought event, although with delay. The tweets also showed a progressive decrease during the drought event to ultimately rise at the end of the event. The drought drivers and impacts represented much fewer tweets. Their F-score was also lower (respectively 74% and 79%). However, the drought driver category did not display any response outside the GRACE-based drought event.