Validation of localised coastal wind forecasts for artisan fishers of southwestern India

In southwestern India, artisan fishing on small boats is a precarious, uncertain and risk-prone job. Fishers are often compelled to fish even in rough weather, especially during the monsoon season. Regular weather forecasts do not always provide enough information to make decisions on whether small-boat crew can safely go to fish in the coastal waters during the monsoon. This study tests and validates whether a localised marine weather information service, using the Weather Research Forecasting (WRF) model, can contribute to safe artisan fishing. (Photo: Max Martin.) The Indian summer monsoon (ISM) during June, July, August and September (JJAS), defined as the seasonal reversal of winds over the South and Southeast Asian region (IMD, 2024), is characterised by distinct spells of rainfall driven by offshore troughs along the west coast of India, movement of monsoon lows and depressions and upper-/mid-tropospheric cyclonic circulations. It is also marked by seasonal movement of the inter-tropical convergence zone (ITCZ) beyond 10°N during boreal summer. The southwestern state of Kerala, where the monsoon winds first strike the peninsular India, has witnessed extreme weather in recent years, along with a marked increase in cyclonic activity in the adjoining eastern Arabian Sea. Kerala suffered high rainfall and flood events in 2018 and 2019. During the 2017 post-monsoon season, Cyclone Ockhi formed in the southwestern Bay of Bengal and traversed 2500km, moving in west and northwest directions before curving back in a north-northeast direction to make landfall in western India. It rapidly intensified into a Very Severe Cyclonic Storm (with surface wind speeds of 64–119 knots) near the coasts of Sri Lanka and south India (Panda et al., 2022), claiming over 365 lives in India, mostly fishers caught out at sea (Rijiju, 2018). While the Ockhi experience highlights the huge risks associated with extreme weather events, day-to-day fishing remains risk-prone too. In 2016–2021, 345 more fishermen died off the shores of Kerala, including 145 in Thiruvananthapuram, the southernmost district (Kumar, 2021). Thiruvananthapuram is home to about 50 000 artisan fishers, practicing fishing mostly on small boats of 30–34 feet in length and even smaller canoes and rafts engaged in hook-and-line, shore seine, boat seine or net fishing 0–50km from the shore. These small craft are vulnerable to winds over 22 knots, denoted as a strong breeze on the Beaufort Scale, with resulting waves of over 3m in height (WMO, 1998; Martin et al., 2022). The Arabian Sea, especially its eastern part adjoining Kerala, appears to be enduring a recent trend of more frequent and intense tropical storms (Murakami et al., 2017; Panickal et al., 2022; Nirmal et al., 2023). Extreme, uncertain and changing weather and climate affect artisanal fishers, considering their dependence on ocean resources as well as exposure and sensitivity to marine weather impacts (Farquhar et al., 2022). Adverse weather is a well-recognised risk factor in fishing, and marine forecasts play a key role in reducing these risks (Finnis et al., 2019; Finnis and Reid-Musson, 2022). Marine weather forecasts have significantly improved over the past decade. Despite this, publicly available models that accurately reflect localised weather impacts relevant for fishers who work close to the shore are still largely inadequate as our interactions with the local fishers show. The local fishers prefer to work whenever there is a reasonable chance of a good catch even if the sea is rough and risk-ridden, as our studies here show. They do not necessarily pay heed to forecasters' advice to stay away from the sea to avoid risk, but instead look at more detailed, nuanced and timely weather information to help them negotiate risks during the monsoon season. The season is marked by high wind and waves but also prospects of a good catch (Martin et al., 2022). Fishers' bulletins that go as part of the daily weather forecasts of the India Meteorology Department are largely generic. Often there is high wind alert for the entire 580km coast of Kerala state. Storm tracks and maps are given for the entire south-eastern Arabian Sea, not the local shores. Fishers have been demanding and trying to access localised and more detailed forecasts as our local pilot studies and technology tests have shown (Osella et al., 2019; Martin et al., 2020; RGS, 2020). It is evident from the above studies that there is a need for localised and detailed weather forecasting which will benefit the artisan fishers' decision-making. To this end, we use the Weather Research and Forecasting (WRF) model (Skamarock et al., 2021) to predict near-surface wind at high resolutions (Jiménez and Dudhia, 2013). We deploy the WRF model to dynamically downscale global model data to represent the mesoscale flow features on the southwest Indian coast. While there are similar studies using WRF model in storm forecasts over different regions (Mohanty et al., 2010; Nadimpalli et al., 2019; Khaira and Astitha, 2023), such studies directly relevant to the Indian Ocean are rare (Alfahmi et al., 2021). The novelty of this study lies in testing a WRF model forecasts for small boat fishers of south India based on local demand. In forecast verification, we assess forecast quality by comparing forecasts with reanalysis for a good estimate of 'truth' (Dorninger et al., 2018). The 'truth' can be ascertained in situ or ex situ. The metrics used comprise consistency, quality and value. Consistency refers to the match of the forecast to the forecaster's best judgement, the quality to the degree to which the forecast agrees with what happened and the value of the forecast's contribution to beneficial decision-making by the users. Quality is associated with verification by nine attributes – bias, association, accuracy, skill, reliability, resolution, sharpness, discrimination and uncertainty (Murphy, 1993). Such verification helps the service provider monitor, improve and compare the forecast quality of different systems (Murphy, 1993; Wilks, 2011). In this study, the forecasts we test are limited to continuous and dichotomous (Yes–No) categories. They are verified using a combination of root mean squared error (RMSE) calculations, contingency tables and related categorical statistics of skill scores. This paper is organised as follows. In the data and methodology section, we present a description of the data used, the WRF model experiment and the forecast verification methods. Data analysis and results are given in the findings section. Finally, discussions and conclusions are made at the end of the paper. The model is configured on a parent domain bounded at 64°E and 97.05°E, and 0.53°N and 31°N around the Indian subcontinent with 230 × 236 grid points at 0.15° resolution. It has an inner (child) domain limited to the southwest region of India including Kerala and parts of Tamil Nadu, bounded by 73°E and 78°E and 7°N and 16.79°N. It consists of 216 × 117 grid points at 0.05°resolution with 33 vertical levels in which 11 levels are below 2km. The mesoscale, non-hydrostatic WRF model, version 4.0 (Skamarock et al., 2021) was used to produce dynamically downscaled forecasts of 10m wind for the domains of study. The cyclic model simulations were set up and initialised at 0000 utc every day and then integrated for 3 days. Forecasts were made every day for the Indian summer monsoon period (June to September) of 122 days for each year from 2020 to 2022, resulting in 366 simulations. The initial and boundary conditions employed in the cyclic model experiments were obtained from the analysis data of the National Centres for Environmental Prediction (NCEP) Global Forecast System (GFS) (https://rda.ucar.edu/datasets/ds084.6/). A two-way nesting was used at a 3:1 parent-to-child domain spatial ratio, allowing communication between the parent (lower resolution – 15km) and child (higher resolution – 5km) domains (Figure 1). Table 1 denotes the model configurations. ERA5 is the fifth-generation European Centre for Medium-Range Weather Forecast (ECMWF) reanalysis for the global climate and weather. The 4D-Var data assimilation and model forecasts are used to combine archival observations into global estimates to eventually produce ERA5 data (Hersbach et al., 2020). The model has a horizontal grid resolution of 31km (0.28°) and 137 vertical model levels. It provides hourly estimates of multiple atmospheric, land and oceanic climate variables. Atmospheric data are interpolated to 37 pressure, 16 potential temperature and 1 potential vorticity level(s) by a powerful post-processing package (Hersbach et al., 2020). Surface or single-level data are also available for selected metrological variables such as wind speed, temperature, precipitation and top-of-atmosphere radiation. In addition to atmospheric and surface data, ERA5 includes information about uncertainties for all variables at reduced spatial and temporal resolutions. Consistent with WRF forecasts, here we use 10m ERA5 wind speed for our analysis. The WRF model forecasts are compared and verified against corresponding ERA5 reanalysis data sets. Different measures of accuracy and skill scores (Wilks, 2011) are used to assess the forecasts. For continuously distributed data, root mean square error (RMSE) is used to compare the WRF model forecasts with observations. For dichotomous (Yes–No) forecasts, we start by constructing a contingency table (Table 2). From the above table, the skill scores of Accuracy, Bias, Probability of Detection (POD), False Alarm Ratio (FAR), Probability of False Detection (POFD), Success Ratio (SR), Threat Score (Critical Success Index – CSI), Gilbert Score (GSS), Peirce's score (PSS) and Heidke score (HSS) are calculated. Verification is essentially a multi-dimensional problem. For example, to get the overall performance of the dichotomous forecast, POD, SR (1 – FAR) and CSI should be used in conjunction. These equations are mathematically related and can be depicted in a single diagram representing the geometric relationships among them. For detailed analysis, we chose 100km of coast in the study area – a shared cultural and working space of the local artisan fishers – into three sub-regions (Figure 2) with distinct wind characteristics using the following onshore reference points as sub-regional boundaries (from north to south): the lighthouses of Thangasseri (08°52′50″N, 76°33′57″E), Anjengo (08°40.4′N, 76°48.5′E) and Vizhinjam (8°22'59.0"N, 76°58'48.0"E), and Enayam beach (8°13′N, 77°11′E). Regions 1 and 3 include the southern parts of Kollam and the northern parts of Kanyakumari districts respectively. The land part of the grids was masked out. The choice of the regions is based on the intricate wind patterns in these coastal areas. These patterns are influenced by the southerlies from the open ocean, the funnelling effect caused by the Palk Straits and the coastal morphology. The southeast Indian and the northern Sri Lankan terrains on either side of the Palk Straits accelerate wind giving a straight southwestward flow (Gupta et al., 2017). Situated on the southern tip of India, Kanyakumari receives unhindered air flows from the open ocean around it during the summer monsoon as well as during the winter monsoon (October–December), making it India's top wind farm location as shown in Figure S1 (Vachaparambil et al., 2014). An analysis of the frequency distribution of 10m wind shows (Figure 3) that Region 1 has the lowest wind speed with a most prevalent range of 5.4–10.8 knots (10–20kmh−1). In Region 2, the prevalent wind speed is 10.8–13.5 knots and in Region 3, there is a nearly equal distribution of 10.8–13.5 knots and 13.5–16.2 knots wind, making it the windiest place among the three regions. The time series (Figure 3) shows that forecast accuracy is comparatively lower in Region 1 with the highest RMSE and lowest correlation values. In Regions 2–3, the RMSE and correlation values are comparable, with the least RMSE and the highest correlation, showing better performance of the model here. The wind roses (Figure 4) show that the winds during the monsoon months are largely north-westerly and light in Region 1 and moderate in Regions 2 and 3. Region 3 shows a comparatively higher occurrence of strong wind with better model performance. The spatial extent of the wind simulated by the WRF model is notably shorter, particularly in Regions 1 and 2. This can be due to the inherent bias in the model's ability to accurately simulate a wide range of wind directions. The contingency table for Regions 1–3 during 2020–2022 (Table 3) shows that the wind speed increases towards the southernmost region. Skill scores of Regions 1–3 for 2020–2022 show high forecast accuracy for the strong wind category in Region 1 (Table 4). This metric is due to the high number of correct negatives, denoting a low prevalence of such wind. POD shows a high value for the light wind category in Region 1, showing good skill, though with high POFD undermining the skill score to some extent. The model is very good at predicting moderate wind, with 1-day lead forecasts and Region 3 shows the best results. The absence of values in some skill scores is explained by the lack of hits and false alarms for those particular wind categories. As Figure 5 shows, Region 3 has the highest values in GSS, especially with high wind, with a decrease in skill as the lead time increases. The figure shows PSS, HSS and POD and SR values below confirm the same results. Overall scores for Region 2 are good with values of 0.4 or above. However, Region 1 did not perform as well as the other regions. Pearson skill score–Heidke skill score combination for Regions 1–3 (Figure 6) shows that 1-day forecast of moderate and strong wind shows good scores in Regions 2–3. However, Region 1 does not show good performance. As the Probability of Detection–Success Ratio combination for Regions 1–3 (Figure 7) shows, in Region 1 the light wind forecast had a lower bias, though with a slight overestimation. Due to more misses in the moderate wind category, the POD is lower, and SR is greater with a large underestimation of moderate wind speeds. Over region 2, low wind was forecast better, closely followed by moderate wind with a slightly higher bias. Moreover, 1-day lead prediction of the strong wind category is useful with high POD, but with low SR possibly because of overestimation in this category. For Region 3, the prediction of the moderate wind category performed better and is close to the perfect forecast with low bias. The prediction of strong wind during the 2021 and 2022 monsoon seasons was close to perfect. The 10m wind forecasts for the monsoon seasons of 2020–2022 were simulated with 1–3-day forecast lead times. The model was validated using 10m wind ERA5 reanalysis data. The frequency distribution and wind rose figures suggest that most of the wind during the monsoon season blows from the northwest and is of moderate strength. Even though there are changes in wind characteristics across the region over time, the model was able to predict the wind speed and direction with reasonable accuracy. Overall, the results indicate that the model can produce accurate forecasts for wind direction and speed, with particular robustness for moderate wind events at 1–day lead across the study area, followed by 2-day lead. The wind speed in the three regions shows different characteristics. Region 1 experiences less wind speed compared with Regions 2–3, with Region 3 showing the highest wind speeds. The highest model accuracy was observed in Region 3, the windiest part of the study area (Figure 8). The analysis suggests that the model is robust and locally relevant for small-grid forecasts. It can predict moderate wind across the study area with high accuracy. One-day forecasts with the highest accuracy are very useful for the local fishers as most of them fish on day or night trips and they prefer to get daily forecasts before leaving for work. In a scenario of frequent high wind events in Region 3, the model performed well. The wind speeds of this area comprising southern Thiruvananthapuram and the adjoining northern parts of Kanyakumari district record considerably higher wind measurements than the other parts of the study area. The literature shows that Kanyakumari has exceptionally high wind speeds, with an overall mean wind speed of over 17.3 knots (Boopathi et al., 2021). This wind speed, however, is not usually included in routine forecasts meant for Kerala fishers. The model performance here indicates its usefulness in a windy region that requires better attention from the forecasters for safe fishing. While the model is useful for the day-to-day use of fishers, its performance in general could be improved for light and strong wind events. This bias of the model in capturing the high wind and low wind categories as shown in Figure S2 can be due to various reasons like small errors in initial condition, inherent model error, limitation of representation of atmospheric processes in the model and a lack of in situ wind observations along the coast and optimum resolution. This underestimation of the forecasting system during high wind conditions is reported in Martin et al. (2022). Further tests may be required to check whether infrequent high wind events are accurately reflected in the model. Further, better performance of 2- and 3-day forecasts can be useful for fishers of southern Thiruvananthapuram who go on multi-day fishing trips. The validation process in this study is based on the reanalyses averaged for 3h and a 25 × 25km grid size. Further validation with in situ onshore and offshore wind measurements may be recommended. For this, there is a need for in situ observations along the coasts and coastal waters of the study regions that could also be fed into the forecasts. Artisanal fishers find value in localised forecasts along with or often even more than the generic formal forecasts available to them (See Martin et al., 2022). As informed users of forecasts from multiple sources, fishers in general seek localised forecasts to complement traditional forecasts in decision-making (Finnis et al., 2019). These localised forecasts can potentially help them optimally use the number of fishing days, especially during the monsoon season. Using these forecasts, the fishers can make informed decisions to deploy an array of safety strategies such as staying close to the shore, finding quickly accessible and safe launching and landing spots, or abandoning a fishing trip when the weather turns bad (Figure 9). In short, localised forecasts can complement and add value to formal forecasts. Further, the availability of such forecasts can limit the number of fishing days lost due to blanket warnings prompted by high wind events that may not be relevant to local areas where fishers go (Martin et al., 2022). In conclusion, the findings highlight the heterogeneous wind conditions prevalent on the coast of southwestern India, where artisanal small-boat fishing is a key livelihood activity. There is a marked difference in the wind regimes along this coast, with the wind speed increasing towards the south. This calls for localised, sub-regional and cross-state-border wind forecasts that are highly relevant for artisanal fishers for safe fishing. Such forecasts could be a significant step in the worldwide move of weather services towards impact-based forecasting that shows what the weather does to different categories of fishers in different locations, rather than merely telling the public what the weather, in general, will be (Campbell et al., 2018). Prabhath H. Kurup: Writing—original draft; writing—review & editing; investigation; validation; formal analysis; data curation; visualisation. Sukumarapillai Abhilash: Conceptualisation; methodology; writing—review & editing; formal analysis; writing—original draft; supervision; project administration; software; resources. Sarang Mavath: Writing—review & editing; investigation; validation; data curation. Vadakkethil K. Mini: Writing—review & editing; validation; methodology, resources. Netsanet Alamirew: Writing—review & editing; validation; methodology; resources. Max Martin: Funding acquisition; writing—original draft; writing—review & editing; project administration; conceptualisation. This research was supported by generous grants from the Sussex Sustainability Research Programme, University of Sussex (through the ESRC Impact Acceleration Account) and the UKRI Global Challenges Research Fund (Grant ES/T003103/1). A pilot study that included technology tests was supported by the Royal Geographical Society. We gratefully acknowledge support for this research from the University of Sussex Forecasting with Fisher's project colleagues Dr Yi Wang, Dr Johnson Jament, Dr MS Visakh, Dr Kate Howland and principal investigator Prof Filippo Osella; Dr Baiju KK at the Advanced Centre for Atmospheric Radar Research, Cochin University of Science and Technology; Mr Niyas NT at India Meteorological Department (IMD) Thiruvananthapuram and Dr Balachandran S, head of the office, IMD Regional Meteorological Centre, Chennai. Our sincere thanks to IMD Director General Meteorology Dr Mrutyunjay Mohapatra for his kind support and advice; and to artisan fishers of Thiruvananthapuram for their valuable feedback and suggestions for this study. Empirical data underpinning this research, plus other environmental and contextual information, will be available from the UK Data Service (at https://ukdataservice.ac.uk). Access is subject to registration at the UKDS, and due to ethical considerations, some ethnographic data might be embargoed until 2027. Figure S1. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.