Hydro-climatic Risk Assessment and Communication for Smallholder Farmers in Maharashtra

(1)

IN

DEGREE PROJECT TECHNOLOGY, FIRST CYCLE, 15 CREDITS

STOCKHOLM SWEDEN 2021,

Hydro-climatic Risk Assessment and Communication for

Smallholder Farmers in Maharashtra

ELIN EKSTRÖM JONNA HALONEN

KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF ARCHITECTURE AND THE BUILT ENVIRONMENT

(2)

Supervisors

Luigia Brandimarte

KTH - Resources, Energy and Infrastructure

Christianne Luger

TU Delft - Water Management

Saket Pande

TU Delft - Water Management

(3)

Abstract

Smallholder farmers often have great entrepreneurial qualities that build on generations of experience.

However, many farm management practices are poorly adapted to current climate change conditions.

In order for farmers to understand the risks they are undertaking by following certain farming practices and to adapt accordingly, a decision support tool is being developed by researchers at TU Delft. The tool runs a socio-hydrological model, created in Python, in the back-end and provides farmer specific investment and profit data for different crops in the front-end. The aim of this study is to develop a risk assessment process that integrates hydro-climatic variability in the decision support tool, and to identify ways of communicating risk to smallholder farmers in Maharashtra, India. Two sources of variability were characterised based on a literature review of Indian farmers’ own risk perceptions; the untimely onset of the Indian Summer Monsoon and the frequency of dry spells. A sensitivity analysis was then carried out to investigate their respective effects on the farmers’ crop yields. The method proposed to evaluate these risks used a single variable, precipitation data, and a two-dimensional risk matrix to compound the two risk factors, over a time span of 14 years (2003-2016). However, the results indicate that it might be more beneficial to define dry spells in terms of crop water stress, instead of a precipitation threshold. This study also proposed a method for translating a cumulative distribution curve into a risk representation that is adapted for low-literacy users by combining numbers and text with graphics, color and voice descriptions. Ultimately, however, the usability of the tool cannot be determined solely through literature, but must involve the end-users in its design.

(4)

Sammanfattning

Småskaliga jordbrukare är goda entreprenörer som samlat på sig kunskaper och erfarenheter över flera generationer. Däremot är vissa metoder som jordbrukarna använder sig av idag för att förvalta sitt jordbruk inte anpassade till nutida klimatförändringar. För att jordbrukarna ska förstå riskerna som de åtar sig vid valet av dessa metoder försöker forskare vid TU Delft nu ta fram ett verktyg för att underlätta jordbrukares förmåga att ta självständiga men välgrundade beslut om sitt jordbruk.

Verktyget är baserat på en socio-hydrologisk modell som är framtagen i Python och som förser specifika investerings- och inkomstdata för enskilda jordbrukare. Syftet med detta kandidatarbete är att bidra till verktyget genom att undersöka de hydroklimatiska risker som uppstår till följd av föränderliga och osäkra klimatologiska förhållanden för jordbrukare i delstaten Maharashtra, Indien.

Två riskfaktorer karakteriserades baserat på en litteraturstudie om indiska jordbrukares riskuppfattningar: avvikelser i starten på den indiska sommarmonsunen och antal torrperioder under monsunsäsongen. Dessutom utfördes en känslighetsanalys för att undersöka om och hur den existerande modellens utdata av skörd påverkades av de valda riskfaktorerna. Monsunstarten och torrperioderna togs fram genom metoder som enbart använde historiska nederbördsdata över tidsperioden 2003-2016 och kombinerades sedan med hjälp av en tvådimensionell riskmatris.

Resultaten visade att det fanns anledning att ifrågasätta hur torrperioderna definierades och att det kan vara mer fördelaktigt att undersöka vattenbrist för grödan, snarare än att enbart förlita sig på nederbördsdata. Vidare föreslog denna studie en metod för att översätta en kumulativ fördelningsfunktion till en grafisk riskframställning som är anpassad till användare med låg läskunnighet genom att kombinera siffror med text, grafik, färg och ljudförklaringar. I slutändan kan dock inte användbarheten av verktyget enbart avgöras utifrån litteratur, utan måste även inkludera återkoppling från slutanvändarna.

(5)

List of Figures

1.1. The study area in Maharashtra. 7

2.1. Workﬂow diagram. 11

2.2. Example of the calculation of the monsoon onset and demise for a single farmer and year. 13 2.3. Example of combining onset deviation and dry spells in a two-dimensional matrix. 14

3.1. eCDF for a farmer in Pandhari village 17

3.2. eCDF for a farmer in Wasri village 17

3.3. Actual crop evapotranspiration (ACET) and potential crop evapotranspiration (PCET) for

one farmer in Pandhari village. 19

3.4. Actual crop evapotranspiration (ACET) and potential crop evapotranspiration (PCET) for

one farmer in Mowada village. 20

3.5. Common visualisations of uncertainty and probabilities. 22

3.6. Visualisation of the bar chart, with accompanying cumulative distribution. 23

3.7. Chosen HTML colors and their respective HEX and RGB codes. 23

3.8. Risk visualisation option 1.0. Vertical bar chart. 25

3.9. Risk visualisation option 2.1. Horizontal bar chart and vertical legend. 25 3.10. Risk visualisation option 2.2. Horizontal bar chart and horizontal legend. 26

List of Tables

1.1. Summary of literature review on farmers’ risk perceptions. 9

3.1. Two-dimensional risk matrix for a farmer in Wasri village. 17

3.2. Yield for one farmer in Pandhari village during the synthetic year created for 2017, with

diﬀerent shifts in the departure of monsoon. Soil depth: 483 mm. 18 3.3. Yield for one farmer in Mowada village during the synthetic year created for 2017 with

diﬀerent shifts in the departure of monsoon. Soil depth: 56 mm. 18 3.4. Yield for one farmer in Pandhari village during the year 2016 when the number of dry spells

is normal (1, 10, 11 and 12) and when dry spells have been added to week 3, 6 and 8 of the

monsoon season. Soil depth: 483 mm. 19

3.5. Yield for one farmer in Mowada village during the year 2016 when the number of dry spells is normal ( 10, 11 and 12) and when dry spells have been added to week 1, 5 and 8 of the

monsoon season. Soil depth: 56 mm. 20

3.6. Mean probability estimates (%) of probability phrases from literature. 24

(7)

1. Introduction

1.1. Background

Agriculture and climate are inseparably linked. Agriculture relies on natural hydrological processes, which underpin the livelihood of farmers and the overall wellbeing of rural communities (Hiremath &

Shiyani, 2013). Meanwhile, unsustainable farming practices are an important driver of environmental degradation. Between 2007 and 2016 an estimated 23% of total anthropogenic greenhouse gas emissions derived from agriculture, forestry and other land use (IPCC, 2020). In the context of sustainability, farming is thus guided by three interconnected principles of being (i) economically feasible for farmers, (ii) socially equitable to rural communities, and (iii) environmentally friendly (Bisht et al., 2020). These three pillars (society, economy and environment) are a common conception of sustainable development. Although there generally is a need to recognise all three dimensions, there can be situations which require a subordination of some sort (Purvis, Mao & Robinson, 2018).

For instance, meeting essential human needs may be prioritised over environmental considerations.

That has been the premise of this thesis and hereafter, the sustainability term refers primarily to the social and economic dimensions.

1.1.1. Problem Outline

Hydro-climatic variability imposes significant constraints to sustainable farming practices, primarily through its adverse impact on crop production. One region that has been particularly affected by such variability is the Indian state of Maharashtra (Swami et al., 2018). Maharashtra accounts for about a quarter of India’s drought-prone districts and 73% of its geographical area is classified as semiarid (World Bank, 2008). The same region is being studied in an ongoing research project conducted at Delft University of Technology (TU Delft). Researchers aim to develop a decision support tool that can help smallholder farmers in Maharashtra to understand the risks associated with certain farming practices, and ways to adapt to those risks. The tool, called Makara, is a mobile application that runs a socio-hydrological model in the back-end and that provides farmer specific investment and profit data for different crops in the front-end (for more information about the project, readers are encouraged to visit the project website at www.makarainit.com). However, further research is needed to translate the model outputs into quantified risk, in order to include the uncertainty in crop yield due to hydro-climatic variability. The risk should be communicated to the farmers in a manner that is easily comprehensible and that fits their current decision-making process and literacy level. In collaboration with the research team at TU Delft, this study aims to develop such a risk assessment and communication process.

Although hydro-climatic variability can manifest itself in different ways, monsoon variability is often highlighted in the Indian context (Swami et al., 2018, 2021). Several studies suggest that the Indian Summer Monsoon (ISM) plays an important part in crop production and equally marks the most critical event in the economic calendar of rural India (World Bank, 2008; Swami et al., 2018; Misra et al., 2017; Prasanna, 2014). Over 70% of the annual precipitation is received during the monsoon season (June-September), making this period highly influential in decisions regarding crop choice, sowing and harvesting time as well as irrigation (World bank, 2008; Swami et al., 2018). Considering

(8)

that Indian agriculture contributes to around 43% of the country’s total employment and 17% to GDP, monsoon variability needs immediate attention from scientists and policymakers (World Bank, 2019;

Ministry of Finance, 2021). So far, this attention has largely been directed towards wide geographical scales and most often to India as a whole. However, according to Swami et al. (2018), formulating climate-related policies for these large domains may lead to under-representation of regional problems. The World Bank (2008) highlights India's vast geographic diversity and that climate projections are likely to be varied, with certain regions experiencing increasing rainfall and floods, while others may be subjected to less precipitation and longer dry spells. Still, little is known about the impacts of climate variability on the meso (districts) and micro (villages, individual farmers) levels (ibid.). The research work carried out in this project aims to fill this gap by specifically addressing hydro-climatic variability for smallholder farmers in Maharashtra and by proposing a methodology tailored to fit local conditions. More specifically, four districts in Maharashtra are covered in this study: Amravati, Wardha, rural Nagpur and Yavatmal (see Figure 1.1).

Figure 1.1. The study area in Maharashtra. Marked are the four districts of interest: Amravati, Nagpur, Wardha and Yavatmal. The scale bar applies to the district's data frame.

Before further narrowing the scope of this study, some comments have to be made on the impacts of climate change on farmers. Since India is characterised by substantial differences in social, demographic and economic conditions, not all farmers experience the same effects (Shukla et al., 2019). Several studies suggest that smallholder farmers are more vulnerable to hardships from climate change than high-resource-endowed farmers. Impacts may range from decrease in income to occupation shifts, migration and suicide (Shukla et al., 2019; World Bank, 2008; Hiremath & Shiyani, 2013; Pande & Savenije, 2016). In fact, the Maharashtra region has recently witnessed many suicides of farmers susceptible to debt, lack of irrigation, fluctuating commodity prices and climate variability (Pande & Savenije, 2016), which is a clear indicator that the third Sustainable Development Goal (SDG) on good health and well being is far from being met (United Nations, 2017). Although there is no universally accepted definition of a smallholder, farmers of this type usually derive their livelihood

(9)

from landholdings less than 2 to 5 hectare and about 10-20 heads of livestock (Narayanan and Gulati, 2002). However, it is common for smallholders to own less than 0.2 hectare of land and only 2 or 3 of the latter (ibid.). In order for climate change adaptation policies to be equitable and effective for all farmer types, more targeted efforts are called for. Therefore, this study specifically addresses the needs of smallholder farmers. By building resilience of those in vulnerable situations and reducing their exposure to climate-related hazards, important steps are taken towards fulfilling SDG1 (ending poverty) and SDG13 (combating climate change and its impacts) (United Nations, 2017).

1.1.2. Conceptualisation of Risk

So far, most of the attention in this introduction has been directed towards hydro-climatic variability, which indeed is the main source of uncertainty for smallholder farmers (Pande & Savenije, 2016).

Here, another term is introduced and that is the concept of risk. This thesis, like many other studies, makes a distinction between the two terms. While risk can be estimated from prior information or projected data from a weather generator, uncertainty may denote a lack of knowledge or an inherent variability of the system studied (Selvaraju, 2012; Walker et al., 2003; Sparks et al., 2018). The focus of this study is on the former since it will use historical rainfall data to estimate future crop yield outcomes. The uncertainty, stemming from situational hydro-climatic variability, and the resulting impact (i.e. effect on yield) is however what generates this risk. In this study, risk is defined as the probability of a bad outcome (Hardaker, 2000). In fact, modelling the concept as a function of probability and consequence is a widespread approach (Walker et al., 2003). More specifically, this study follows the definition of agricultural risk proposed by Selvaraju from the FAO (2012, 71) as;

“the probability of a defined hydro-meteorological hazard affecting the livelihood of farmers [...]”.

Here, livelihood refers to crop yield and the choice of these risks was based on a literature review, which is summarised in Table 1.1.

To ensure the success of the risk assessment it is necessary to consider the farmers’ own risk perceptions. Several studies suggest that they strongly affect how the farmers understand and deal with climate induced risks and undertake specific adaptation strategies (Raghuvanshi & Ansari, 2019;

Singh et al., 2020; Banerjee, 2014; Ansari et al., 2018). For this reason, a preparatory literature review was conducted to identify the most commonly perceived hydro-climatic risks (see Table 1.1).

Banerjee (2014) remarks that few studies on perception have been conducted in the drylands of India and to the authors’ knowledge only two such studies exist for Maharashtra. Therefore, the literature review includes studies from other Indian states. As previously mentioned, climate change is location-specific. However, the review aims primarily to identify types of risk, which are more or less consistent across India (e.g. the onset of the monsoon, frequency/duration of droughts or floods, temperature), and not their magnitude (e.g. early/late, increase/decrease, high/low). Finally, it is worth mentioning that farmers’ risk perceptions seem to coincide well, and sometimes in almost complete accuracy, with historical meteorological data and observed climate trends, which is why studies of this kind were omitted in the review (Banerjee, 2014; Shukla et al., 2019; Hein et al., 2019). Two studies also suggest that farmers’ perceptions on types of climatic risks (N.B. not impacts) differ neither within farmers of the same farm size, nor between farmers of different landholdings (Banerjee, 2014;

Shukla et al., 2019). It is therefore assumed that most smallholder farmers hold similar views.

(10)

Table 1.1. Summary of literature review on farmers’ risk perceptions

Reference Geographic focus Method Key ﬁndings

Ansari et al. (2018) Uttar Pradesh, India Interviews with 120 farmers Increased duration of dry spells, temperature and frequency of heavy rains along with untimely onsets of the rainy season were the four most common concerns

Banerjee (2014) Maharashtra and

Andhra Pradesh, India Group discussions and individual interviews with 273 farmers

Farmers perceived an increase in

temperature, higher variability in quantity and distribution of rainfall as well as late onset of rains

Dhaka (2010) India Interviews with 500 farmers Many farmers believe precipitation has declined and temperatures have increased along with later onsets and earlier demises of the monsoon with longer dry spells

Kelkar et al. (2008) Uttarakhand, India Interviews with 62 farming

households Almost all the households noticed that they could no longer rely on the timely onset of the monsoon and that rainfall had declined in quantity

Kumar & Radha

(2016) Andhra Pradesh, India Interview with 504 farmers The farmers perceived the delay in monsoon and periodic deﬁcit as the largest threats to crop production

Raghuvanshi &

Ansari (2019) Uttarakhand, India Interviews with 30 farmers The frequency and extent of dry spells and soil erosion due to heavy rainfall were two of the greatest concerns

Shukla et al. (2019) Uttarakhand, India Interviews with 241 farming households across ﬁve diﬀerent farmer types

Respondents across all ﬁve farmers types observed an increase in summer temperature and decrease in summer and winter

precipitation Singh et al. (2020) Uttar Pradesh, India Interviews with 60

smallholder farmers Farmers perceived that erratic rainfall and extreme events (rainstorm, ﬂash ﬂood, longer dry spells) have increased

Suresh et al. (2017) Maharashtra, India Field survey with 244

farmers Farmers perceived the late onset of the monsoon as the major risk and low rainfall or drought as the second most important risk Varadan & Kumar

(2014) Tamil Nadu, India Survey with 200 farmers The most common observations were that the quantity of rainfall has decreased over the years, and that the onset of the monsoon is more delayed today than before

According to the literature review, the most commonly perceived hydro-climatic risks are the onset of the monsoon season, amount of rainfall and dry spells. Since quantity of rainfall is included in the very definition of dry spells, which will be discussed further in the methodology, the choice was made to only look at the timing of the monsoon and the frequency of intra-seasonal dry spells. Although the literature suggests that these risks play an important part in crop production, it is unknown how the existing socio-hydrological model and local conditions respond to them. Therefore, a sensitivity analysis will be carried out in order to determine how the target output variable, crop yield, is affected by isolated, risk-related changes in input data.

(11)

1.2. Aim

The aim of this project is to empower smallholder farmers in Maharashtra by enabling them to make informed decisions regarding crop choice, based on predictions of their expected crop yields. The process of farmer empowerment will consist of raising awareness of the uncertainty in yield output due to hydro-climatic variability, and giving access to a risk assessment adapted to their needs. The hope is that this will enable the farmers to better manage and to self-assess their abilities to adapt to defined hydro-climatic risks in the future.

The research questions are threefold and involve both the assessment and communication of risk:

1. How can the two sets of risk, monsoon onset and dry spells, be assessed and incorporated in the decision support tool?

2. How is crop yield affected by a sensitivity analysis designed to evaluate the chosen risk factors?

3. How can these risks be effectively communicated to the farmers in a user friendly manner, taking into consideration their literacy level?

1.3. Objectives

By writing code in Python, this project will develop a risk assessment process that integrates the two hydro-climatic risks of a smallholder farmer and through this method transform socio-hydrological model outputs into appropriate quantified risk factors. The manner of how this risk is presented to farmers through a decision support tool is also investigated, as it should be easily comprehensible and fit their current decision-making process.

(12)

2. Methodology

The methodology is divided into two main parts: section 2.1 introduces the risk assessment process, including data and definitions and section 2.2 describes the methodology for communicating these risks to the farmers in the study area. Figure 2.1 gives an overview of the workflow during this project.

Figure 2.1. Workﬂow diagram. Only the green boxes are treated in this study. That is, the prototype design and implementation of the risk representation as well as feedback from the farmers is outside the scope of this study.

2.1. Risk Assessment

2.1.1. Socio-hydrological Model and Data

The socio-hydrological model used in this study originates from the modelling framework proposed by Pande and Savenije (2016), and has been realised using the Python programming language by students at TU Delft. The model incorporates five main variables: water storage, household capital, livestock, fodder and soil fertility, who interact with one another through a series of feedback mechanisms. Gridded daily evaporation and precipitation data from 2003 to 2016 were used to simulate the crop yield and income for each of the farmers for the given years. The methodology proposed in this study will however be adapted to a 47 year precipitation record before being incorporated in the decision support tool. Furthermore, the historical data was used to predict the farmers’ crop yield and income for one season ahead. The precipitation data has been obtained from the Indian Meteorological Department (IMD) and has a grid spacing of 0.25° x 0.25°. Moreover, the

(13)

model runs on the micro-level scale for 308 different farmers in the four districts of Maharashtra, which have been previously identified through a survey conducted by TU Delft (Pande et al., 2019).

However, the validation of the model at the farm level is poor and Djohan (2021) concludes that the model displays inaccuracies between the predicted yield for individual farmers and their observed yield. In this study, the risk assessment process was performed by defining and calculating the risk of an untimely monsoon and the frequency of dry spells. The existing data and model were used to obtain each farmer’s crop yield, which was the main output variable used to calculate the risk. For the sensitivity analysis, the precipitation record was manipulated to create synthetic data, which is explained in more detail in section 2.1.5. below. All calculations were made as an extension to the already existing Python code.

2.1.2. Deﬁnition and Calculation of Risks

Onset and Demise Dates of the Indian Summer Monsoon (ISM)

The local onset and demise of the ISM was defined and calculated using only precipitation data, a method first proposed by Liebmann and Marengo (2001) and that has since been adapted by many others (Bombardi et al., 2017; Bombardi et al., 2019; Noska & Misra, 2016; Karmakar & Misra, 2019). The demise date of the ISM was not part of the risk assessment, but only served to determine the length of the monsoon season for the dry spell calculations. By using only one variable, the method has the advantage of simplicity, while still showing consistency withthe seasonal evolution of other dynamic and thermodynamic variables associated with the ISM (Noska & Misra, 2016).This study specifically followed the methodology proposed by Bombardi et al. (2017) and looked at the interannual variability of the monsoon onset during the years 2003-2016. Although the authors made a distinction between the monsoon, which indeed is a complex large-scale phenomenon, and the rainy season, which can be said to accompany a particular monsoon system, this study chose to maintain the monsoon terminology. Apart from other studies having done the same (Noska & Misra, 2016;

Karmakar & Misra, 2019), it was hypothesized that this term is more established with farmers.

The onset and demise of the ISM for each farmer coordinate was based on accumulated precipitation anomalies S from Eq. (1) and is illustrated in Figure 2.2.

𝑆 =

(1)

𝑖 = 𝑡

0

𝑛

∑ (𝑃(𝑖) − 𝑃

𝑐)

P(i) is the daily precipitation at day i and P_c is the annual daily average precipitation. The calculation starts on day t₀ and is the same for every historical year. The choice of t₀ depends on the region of interest and should be chosen within the dry season for the onset and within the wet season for the demise. For the onset, Bombardi et al. (2017) proposed a date in early April for the Indian subcontinent, by referencing the IMD’s official onset date of the rainy season as May 10th. However, this would generate too many false onsets in the given case, since according to the IMD (n.d.) the local monsoon onset around Maharashtra is not until 10-15 June. Moreover, Adamson and Nash (2013), who studied the long-term variability in the monsoon onset over the Mumbai area in Maharashtra, found that the earliest identified date between 1781 and 2011 was 23 May. For these reasons, t₀ for the onset calculation was set to the 20th of May (140 in Julian days). Choosing t₀ for

(14)

the demise calculation was proven more challenging. Syroka and Toumi (2004) found that the withdrawal of the ISM was more variable than the onset, which was also reflected in the 2020 end of season monsoon report by the IMD (n.d.). The normal demise date for Maharashtra was said to be around 5-10 October but the monsoon did not start to withdraw until the 26-27th (ibid.). In this study, t₀ for the demise calculation was set to 27th of September (270 in Julian days), which accommodates more to the “normal” demise date proposed by the IMD.

Figure 2.2. Example of the calculation of the monsoon onset and demise for a single farmer and year. The left axis shows observed precipitation values and the right axis shows the values of S, S smoothed (Ss), and the ﬁrst derivative of Ss, the latter has been scaled ﬁve times for clarity.

Since the beginning of the year is during the dry season, S (pink line) will initially assume negative values. Once the monsoon season starts, there will be an inflection in S. In order to avoid false onsets or demises, the S curve was smoothed (blue line, Ss) using a 3-point moving average and the first derivative of the smoothed curve (orange line, d(Ss)/dt) was taken. Then, starting the calculation from t₀(left black line), the first day when the derivative changed from negative to positive values was considered the onset of the monsoon (green line), as long as the positive values persisted for 3 consecutive days. Similarly, the first day when the derivative crossed from positive to negative values, and they persisted for 3 days, was considered the demise of the monsoon season (red line). The onset and demise was found for all 14 simulation years and a mean onset date was calculated thereafter. The risk was expected to arise as the onsets deviate from this mean. In other words, the premise was that the larger the departure from this mean onset date, the greater the risk of affecting the crop yield. A more exact correlation between this deviation and its effect on crop yield was attempted to be found through a sensitivity analysis, which is described in section 2.1.6.

(15)

Frequency of Dry Spells

There is no single, universal definition of a dry spell. Swami et al. (2018) have looked at definitions proposed by various researchers. One study defined a prolonged dry spell as rainfall below 2.5 mm/day for more than 4 days (Dash et al., 2009). Another analysed the duration between 2 subsequent wet spells (Singh & Ranade, 2009) and a third used standard deviation from daily precipitation for more than 3 consecutive days (Singh et al, 2014). According to Barron et al. (2003), definitions that are solely based on rainfall characteristics are called meteorological dry spell assessments. In order to maintain a uniform risk assessment process between the onset and dry spell calculations, the choice was made to use such a definition. Alternatively, one could employ an agricultural dry spell definition, which considers crop water stress (i.e. deficit of crop accessible soil water), and which Barron et al. (2003) argues is often of more value to the farmer. The latter would require a water balance equation and data such as crop water demand, water holding capacity of the soil and crop water uptake capacity. In this study, the choice was made to define a dry spell as an extended period of dry days, where a dry day is a day with precipitation less than a preselected threshold. By adopting the same approach as Kumar et al. (2019), the threshold was based on the IMD’s definition of a rainy day, which is 2.5 mm/day, and the time period was set to 1 week. In other words, the frequency of dry weeks was calculated for each of the 14 simulation years, under the constraint that precipitation had to be less than 20 mm for a given week during the monsoon season.

The previously found onset and demise dates were used to find the length of the season. This time, it was assumed that a higher frequency of dry spells equated to a higher risk of affecting the crop yield.

2.1.3. Two-dimensional Risk Matrix

In order for the risk visualisation to be as concise as possible for the farmers, the two risks were compounded into one single risk output. This was done by creating a two-dimensional matrix, which could be expanded to a multi-dimensional matrix if more risk factors were to be included. In each year of the historical data, the onset for the rainy season and number of dry spells were counted. The frequency of the combination of the two hydro-climatic events happening at the same time was then inserted in a matrix, see Figure 2.3 for visual clarification.

Figure 2.3. Example of combining onset deviation and dry spells in a two-dimensional matrix.

(16)

The probability of a combined event was calculated as the frequency of one event divided by the total frequency, Eq. (2).

(2)

𝑃𝑥= _𝐹^𝐹^𝑥

𝑡𝑜𝑡

where, Pxis the probability of one combination, Fxis the frequency of one combination and Ftotis the total frequency of all combinations. Each combined event was then attached to a yield. For example, if the combined event of one dry spell happening together with a two day early onset had a frequency of one and this happened in year 2004, the yield connected with the combined event was the yield of 2004. If on the other hand the combined event happened in 2004 and 2007, the yield output would be the average of these two years. The probability attached to the yield made up the risk. An important assumption was made here; the climate was assumed to be static. Each year of the historical data was therefore assumed to be of equal value, when in reality data from 2016 might have greater value when simulating the future than 2003 for example. Milly et al. (2008) states that stationarity is dead, due to anthropogenic climate change, and should no longer serve as a default assumption in water-resource risk assessment. Besides the length of the historical data, the forecast horizon also influences this assumption. Since this model only predicts one season ahead, it can be assumed that the results would not be affected to a large degree by climate change.

2.1.4. Crop Yield Distribution

A crop yield distribution can be used to model the risk exposure, i.e. to visualise the crop yields attached to their respective probabilities (Hennessy, 2009). Characterising the shape of this distribution has important implications for production decisions under uncertainty and risk-efficient farm planning (Just & Weninger, 1999; Ramirez et al., 2001). Oftentimes, risk is quantified by use of standard deviation, which is a good measure of variability for symmetric, normal distributions.

However, a common and long-standing assertion is that crop yields are skewed and non-normal (Harwood et al., 1999, Hennessy, 2009). In fact, most yield distributions seem to be negatively skewed (i.e. have a longer lower tail), which reflects the fact that biological constraints inherent to the crops and environmental factors (e.g. pest damages and weather) often lead to lower crop yields (Hennessy, 2009; Xiang et al., 2012). Although some studies point out that a consensus regarding nonnormality has not emerged in the agricultural literature (Just & Weninger, 1999; Ramirez et al., 2001), this study chose to assume a non-normal distribution.

Since the probabilities were derived from historical observations, this study used an empirical cumulative distribution function (eCDF), which provides the probability of a data point falling above or below a speciﬁed value (Harwood et al., 1999). The curve was created by first sorting the output yields from the two-dimensional risk matrix in ascending order, and then taking the cumulative sum of their probabilities. By using linear interpolation between the yields and their cumulative sum, the eCDF was plotted. However, it was assumed that this representation of risk might be too complicated for the farmers to understand. This leads into the risk communication part of the project, which is explained in section 2.2.

(17)

2.1.5. Sensitivity Analysis

In order to examine if and how the crop yield was affected by the two chosen risks, a sensitivity analysis was performed. Two farmers were chosen for the analysis and were used for both the monsoon and the dry spells part. These two were taken at random except for the fact that they had a great variety in soil depth, which felt reasonable to vary as to previous arguments that mention that crops' access to soil water might have a great effect on yield outcome. One farmer had a soil depth of 56 mm and was located in Mowada village and the other had a soil depth of 483 mm and was located in Pandhari village. For the monsoon onset, a synthetic year was created by averaging the daily precipitation over the historical years (2003-2016).This year was then placed as the new last year in the historical data, that is as 2017, and the socio-hydrological model was run to determine the yield output. Thereafter, the synthetic year was manipulated by shifting the daily precipitation of the created average year forward or backward by a certain number of days and reexamining the yield. In other words, the precipitation record for the average year was rotated. For example, if the monsoon was shifted one day forward the daily precipitation for the 31st of December was moved to the 1st of January and so on. The magnitude of the shift was determined by looking at the historical deviations of the onset date from the mean onset date. The yields from the different shifts were recorded in a table to see if any variation occurred.

As for the dry spells, there was a concern that creating a synthetic year would cancel out low precipitation days due to averaging. Therefore, the decision was made to examine the current last year in the historical data, 2016. First, the number of dry spells and their placement during the monsoon season was determined. The precipitation for the found dry spells was calculated and the average was taken, this represented an average dry spell. The actual sensitivity analysis was done by scaling the precipitation of a week that was not classified as a dry spell, in the beginning, middle and end of the monsoon season to the average precipitation of the previously determined dry spells. The yield was determined for the three scenarios occurring separately. As previously mentioned, there is a possibility that agricultural dry spells might be of larger importance to the crop yield and the crop water was therefore also examined in the sensitivity analysis. The potential crop evapotranspiration was plotted against the actual crop evapotranspiration to determine if the result found in the yield outcome could be explained by these plots. If the potential and actual crop water overlapped while increasing the number of dry spells, this would result in the yield not being affected since the crops would still have a sufficient amount of water.

2.2. Risk Communication

The risk communication part takes its starting point where the crop yield distribution left off. The main concern here was to translate the distribution curve into a more comprehensible risk representation. To do so, a literature study was conducted to obtain insight in ways of communicating risk, and in user-centred design for illiterate and semi-literate users. The actual transformation of the model outputs to the suggested risk visualisation will be carried out by the front-end team at TU Delft.

At a higher level, this study follows a co-creation methodology, where the initial prototype design, based on the literature review, will be tested with the farmers and they will be able to give feedback on its usability. This part, however, is outside the scope of this study.

(18)

3. Results

3.1. Risk Assessment

3.1.1. Two-dimensional Risk Matrix and Empirical CDF

Table 3.1 shows the two-dimensional risk matrix for a farmer in Wasri village, ordered by ascending yield. Negative departure values from the mean monsoon onset date indicate an earlier date. It can be seen, for instance, that this farmer experienced two years when the monsoon onset was six days late from the average and there were seven dry spells during the monsoon season. There were also two years when the monsoon onset was two days early and there were four dry spells during the monsoon season. All other combinations of events only occurred once. Moreover, no evident correlation can be seen between the two risk factors and the yield outcome for this farmer. For example, one dry spell during the monsoon season is attached to years with both the highest and the second lowest yield.

Table 3.1. Two-dimensional risk matrix for a farmer in Wasri village.

0 1 2 3 4 5 6 7 8 9 10 11

Departure from mean monsoon

onset date [days] 1 -6 3 -3 -17 6 -2 2 11 4 -4 5

Number of dry spells during

monsoon season 6 1 5 5 10 7 4 4 5 3 2 1

Frequency 1 1 1 1 1 2 2 1 1 1 1 1

Average yield [kg/ha] 1438 1643 1678 1734 1740 1743 1784 1790 1856 1984 2056 2081

Probability 0.071 0.071 0.071 0.071 0.071 1.143 1.143 0.071 0.071 0.071 0.071 0.071

Figures 3.1 and 3.2 display the empirical CDFs for two farmers, one in Pandhari village and one in Wasri village. These two were chosen to emphasise that all eCDFs are unique to the farmers and that no two curves look quite the same. The most likely crop yields are found where the eCDF is the steepest. Furthermore, the farmer in Pandhari village can be seen to have experienced most yields over 2075 kg/ha and the farmer in Wasri village can be seen to have experienced most yields over 1650 kg/ha. See Appendix 1 for the Python code written to create the two-dimensional risk matrix and the eCDF.

Figure 3.1. eCDF for a farmer in Pandhari village Figure 3.2. eCDF for a farmer in Wasri village

(19)

3.1.2. Sensitivity Analysis Monsoon Onset

Table 3.2 displays the yield for eight shifts in the onset date for a farmer in Pandhari village. The yield can be seen to start decreasing when the departure is after the average onset date. This farmer had a soil depth of 483 mm.

Table 3.2. Yield for one farmer in Pandhari village during the synthetic year created for 2017, with diﬀerent shifts in the departure of monsoon. Soil depth: 483 mm.

Monsoon shift Yield [kg/ha]

-8 2156.18

-5 2156.18

-4 2156.18

-3 2156.18

-1 2156.18

3 2147.69

4 2142.24

9 2119.45

Table 3.3 displays the yield for nine shifts in the onset date for a farmer in Mowada village. The yield can be seen to increase the later the monsoon onset begins. This farmer had a soil depth of 56 mm.

Table 3.3. Yield for one farmer in Mowada village during the synthetic year created for 2017 with diﬀerent shifts in the departure of monsoon. Soil depth: 56 mm

Monsoon shift Yield [kg/ha]

-7 1492.65

-5 1523.82

-4 1537.93

-3 1552.82

0 1596.98

1 1611.69

3 1637.8

4 1649.35

5 1661.09

10 1722.01

(20)

Dry Spells

Table 3.4 and Figure 3.3 show the sensitivity analysis results of the dry spells for a farmer in Pandhari village. The farmer originally had dry spells for week 1, 10, 11 and 12 of the monsoon season 2016.

The average dry spell was 6 mm. A dry spell was added to week 3, 6 and 8 by scaling down the precipitation to 6 mm for these weeks. No change in yield was detected whilst increasing the number of dry spells (Table 3.4). The actual crop evapotranspiration did not change either (Figure 3.3). This farmer had a soil depth of 483 mm.

Table 3.4. Yield for one farmer in Pandhari village during the year 2016 when the number of dry spells is normal (1, 10, 11 and 12) and when dry spells have been added to week 3, 6 and 8 of the monsoon season. Soil depth: 483 mm.

Dry spells during weeks: Yield [kg/ha]

1, 10, 11, 12 2156,18

1, 3, 10, 11, 12 2156,18

1, 6, 10, 11, 12 2156,18

1, 8, 10, 11, 12 2156,18

Figure 3.3. Actual crop evapotranspiration (ACET) and potential crop evapotranspiration (PCET) for one farmer in Pandhari village. ACET shown for normal precipitation and for precipitation with dry spells added to week 3, 6 and 8 for the monsoon season of 2016. Soil depth: 483 mm

Table 3.5 and Figure 3.4 show the sensitivity analysis results of the dry spells for a farmer in Mowada village. The farmer originally had dry spells for week 10, 11 and 12 of the monsoon season 2016. The average dry spell was 6.6 mm. A dry spell was added to week 1, 5 and 8 by scaling down the precipitation to 6.6 mm for these weeks. No change in yield was detected whilst increasing the number of dry spells (Table 3.5). The actual crop evapotranspiration did not change either (Figure 3.4). This farmer had a soil depth of 56 mm.

(21)

Table 3.5. Yield for one farmer in Mowada village during the year 2016 when the number of dry spells is normal (10, 11 and 12) and when dry spells have been added to week 1, 5 and 8 of the monsoon season. Soil depth: 56 mm.

Dry spells in weeks Yield [kg/ha]

10, 11, 12 1293.23

1, 10, 11, 12 1293.23

5, 10, 11, 12 1293.23

8, 10, 11, 12 1293.23

Figure 3.4. Actual crop evapotranspiration (ACET) and potential crop evapotranspiration (PCET) for one farmer in Mowada village. ACET shown for normal precipitation and for precipitation with dry spells added to week 1, 5 and 8 for the monsoon season of 2016. Soil depth: 56 mm.

3.2. Risk Communication

Many researchers have examined the design of user interfaces (UI) for semi-literate and illiterate users, some of which are targeted directly to farmers in India (Joshi et al., 2008; Rege & Nagarkar, 2010; Medhi-Thies et al., 2015). Likewise, there is a substantial body of work that deals with the visualisation of risk, or uncertainty, and the communication of probabilities (van der Bles et al., 2019;

Spiegelhalter et al., 2011; Visschers et al., 2009). This part of the study aimed to combine these two fields of work and ultimately create a risk visualisation that is user-friendly and adapted to the rural context. Lahr and Kooistra (2010) conclude that the visualisation of environmental risk can be a powerful tool to communicate the outcome of complex risk assessments to decision-makers. Yet many approaches for communicating risk are poorly received by end-users (Olivier et al., 2017), mostly because they fail to involve their target users in the design of such tools (Whitman et al., 2015). The farmers' engagement in the design process has not been considered in this study. Instead, the design choice has rested solely on literature. Finally, it is important to note that, based on the 2011 Census of

(22)

India, the literacy rates of the rural population in Amravati, Wardha, Nagpur and Yavatmal were about 86, 84, 83 and 78 % respectively (Government of India, 2011).

Different design principles of UIs have been proposed for low-literacy users. Many researchers have recognised the value of imagery and graphics (Grisedale et al., 1997; Medhi et al., 2006; Ghosh et al., 2003; Parikh et al., 2003). Others have highlighted the use of audio-visual icons and voice descriptions for supplying information normally provided via text (Medhi et al., 2006; Medhi et al., 2007; Rege & Nagarkar, 2010). In two studies, the use of color was found to be a helpful tool in overcoming the users lack of literacy (Joshi et al., 2008; Parikh et al., 2003). Moreover, Parikh et al.

(2003) found that text, even though users could not fully read it, should be provided in the local language since it gave users a greater sense of familiarity and ownership. Finally, some authors have noted that numeric data (e.g. dates, interest percents, sums etc.) can be acceptable in user interfaces, as many low-literate people can read and understand numerical digits. This is likely due to the exposure to number-based data formats such as bus timetables, calculators, phone numbers, and calendars (Ghosh et al., 2003; Parikh et al., 2003; Medhi et al., 2006).

Numeracy, which can be defined as the ability to process basic numerical concepts, has also been highlighted in the risk communication literature. In fact, this characteristic has a large influence on people’s understanding of probabilistic information and ability to make good decisions (Peters et al., 2006; van der Bles et al., 2019; Spiegelhalter et al., 2011). Before moving forward, however, risk communication first has to be defined. This study used a narrow definition of the term, in that it only examined risk messages and the way farmers may perceive them (i.e. understanding). Broader issues that are sometimes involved in risk communication, such as trust, emotion, policy issues and cultural preferences are not touched upon (van der Bles et al, 2019; Steg & Sievers, 2000; Covello et al., 1986). Some comments will however be made on farmers’ decision-making process, since this is closely related to perceptions of risk and uncertainty. With this said, probabilistic information can be communicated using numerical, verbal, and/or visual (e.g. graphical) formats (Lipkus, 2007). For instance, a farmer can be told that the probability of getting a crop yield below 2000 kg/ha is 20%

(numerical) or that it is unlikely (verbal), or the farmer can be shown a cumulative distribution function (graphical). Earlier, it was hypothesized that a CDF would be too complex for the farmers to understand. Examples (i)-(iii) in Figure 3.5. display some common expressions of uncertainty that may replace, but still give an indication of, the underlying distribution. In these cases, one might assume that the underlying distribution is normal since the width, color and density is centered around a mean value. Examples (iv)-(vi) show common visualisations of probabilities.

(23)

Figure 3.5. Common visualisations of uncertainty and probabilities. Adapted from van der Bles et al. (2019) and Spiegelhalter et al. (2011). Crop icon by Endeavorbd, licensed under CC BY-SA 3.0.

In contrast, the CDF is not centered around a mean. Instead, it displays the probabilities of outcomes falling above or below specific values. According to Harwood et al. (1999), this ability makes the CDF especially useful in “safety-first” analyses, which seem to fit the farmers decision-making process well. In risk management, the safety-first approach implies that a decisionmaker first satisfies a preference for safety (such as minimising the probability of bankruptcy) before addressing other, higher-level goals (such as profit maximisation). This method is particularly applicable to individuals where survival, or the satisfaction of basic needs, is of principal concern (Harwood et al., 1999). Some authors have speculated that smallholder farmers, in particular, would be concerned about the probability of their livelihood falling below a certain “disaster level”, before concentrating on maximising their profits (Shahabuddin & Butterfield, 1986; Umar, 2013; Rosa et al., 2019; Senapati, 2020).

With this in mind, the attention then turned to the framing of the probability. For instance, the farmer can either be told that there is a 40% chance of getting a yield below 2000 kg/ha, or that there is a 60% chance of getting a yield above this value. The framing effect has been a widely studied topic, and a common notion seems to be that the same prospect, framed in terms of success, is perceived as more attractive than when framed in terms of failure (Tversky & Kahneman, 1981; Levin & Gaeth, 1988). In this study, the decision was made to communicate a positive frame, or the probability of exceedance, for two reasons. First, presenting the probability of falling below a certain value may lead farmers to believe that high yields are more likely. Secondly, a positive frame may help the farmers to

(24)

see opportunities instead of misfortune. Next, the choice was made to communicate five different probabilities; 10, 25, 50, 75 and 90 percent, and to use a bar chart, since it is valuable for conveying magnitude (Spiegelhalter et al., 2011). These choices are envisioned in Figure 3.6.

Figure 3.6. Visualisation of the bar chart, with accompanying cumulative distribution.

Color also has an important role in risk communication, as highlighted by Borade et al. (2008). They looked at hazard perception associated with colors and safety words among Indian industry workers.

The results showed that the workers mostly associated red with “Danger” and “Warning”, while green was associated with the words “Go” and “Safe”. The color in the risk presentation therefore had to be chosen with care. As seen in Figure 3.6, the highest yields were attached to the lowest probabilities, which made it hard to motivate a color choice. Instead, a monochromatic scheme was proposed, which at the same time is more beneficial to those with red and green color vision deficiency. The choice of shades was based on the yield magnitudes, where the darkest shades were assigned to the highest yields. The actual HTML colors that were chosen, and their respective color codes, are visualised in Figure 3.7.

Figure 3.7. Chosen HTML colors and their respective HEX and RGB codes. The codes are listed from dark to light.

(25)

Alongside color, it was also appropriate to include numerical and verbal probabilities, since no single representation suits all members of an audience (Spiegelhalter et al., 2011; Visschers et al., 2009). The verbal probabilities were chosen against studies that had examined peoples’ numerical interpretations of verbal probability phrases. Table 3.6. summarises the results of these studies for each of the chosen expressions. It should be noted that there were relatively high standard deviations for most phrases and sometimes large differences between studies. Some studies have indeed questioned the precision of verbal expressions (Spiegelhalter et al., 2011; Visschers et al., 2009), but the hope is that the combination of words and numbers will be able to convey an accurate risk message.

Table 3.6. Mean probability estimates (%) of probability phrases from literature.

Chosen phrases Budescu

& Wallsten (1985)

Brun &

Teigen (1988)

Mauboussin &

Mauboussin (2018)

Sutherland

et al. (1991) Willems et al.

(2020)

US Intelligence Community Directive 203 (Wintle et al, 2019)

IPCC guidelines (Budescu et al., 2014)

10 Very unlikely 11 5-20 < 10

25 Unlikely 20 ~ 10-25 29.4 16 20-45 < 33

50 Possible 38 38 ~ 50 40.6 47

75 Likely 74 67 ~ 75 70.1 75 55-80 > 66

90 Very likely 86 80-95 > 90

Finally, following the recommendations from the literature on UIs for illiterate and semi-literate users, the risk visualisation also incorporated graphics. Two icons were meant to give context to the risk assessment and complement a short descriptive text. A rain cloud with an overlying clock and a crossed out water droplet was intended to represent the two hydro-climatic risk factors of an untimely monsoon and dry spells. A plant icon was added to signify that the figure displays crop yield and not say, a weather forecast. It was also recommended that the user interface offers a voice description of the context and content, in order to accomodate to farmers who lack literacy. With all of these factors considered, three options of the risk visualisation have been proposed. One option is to display the bar chart vertically (see Figure 3.8).

(26)

Figure 3.8. Risk visualisation option 1.0. Vertical bar chart. The icons are referenced under section 7.

The two other options contain a horizontal bar chart but different legends, one of which is vertical (Figure 3.9).

Figure 3.9. Risk visualisation option 2.1. Horizontal bar chart and vertical legend.

(27)

The other contains a horizontal legend (Figure 3.10). The choice between these three risk visualisations is discussed under section 4.2.

Figure 3.10. Risk visualisation option 2.2. Horizontal bar chart and horizontal legend.

(28)

4. Discussion

4.1. Risk Assessment

This study has demonstrated that a two-dimensional risk matrix could be used to compound the risk of an untimely monsoon onset, together with the number of dry spells. This method of assessing multiple risks has the benefit of adaptability, in that it is able to include more risk factors if required, and the advantage of simplicity since only one variable, precipitation data, is used. However, it should be mentioned that the dry spell risk is only characterised by the frequency of dry spells and not their placement during the monsoon season, which is a simplification that might affect the result. It is possible that the nature and/or severity of the dry spell risk varies depending on the distribution of the dry spells during the different growth stages of the crop. For instance, early season dry spells might lead to seedling death, while terminal dry spells might cause inadequate crop maturity. As pointed out in the preparatory literature review on farmers’ risk perceptions (Table 1.1), farmers might also be concerned about the duration of dry spells. This study chose to define a dry spell as a fixed week with precipitation below 20 mm and it therefore neglects this temporal aspect.

Another limitation that might be seen with this method is the assumption of static climate. When equal value is given to the years in the beginning and the end of the historical data, the variation of long term climate is neglected. This assumption might be valid for studies that use short precipitation records and short forecast horizons, as long term climate changes might not be seen over short timespans. For long time frames, however, there is reason to investigate if climate change trends can be included. Trends might also be considered when choosing the start date, t0, for the onset and demise calculation. In this study, t₀ for the onset was adjusted after the earliest identified date since 1781, provided by Adamson and Nash (2013), while t₀for the demise accommodated to the “normal”

long-term average proposed by the IMD. Under the assumption of a non-static climate, different start dates could be used for different years and in doing so, the chances of getting false onset and demise dates might be minimised.

Finally, looking back at Table 3.1, no evident correlation can be seen between, for example, lower crop yields and higher frequencies of dry spells or delayed monsoon onsets for the investigated farmer. This might be due to the fact that some risk factors cancel out each other. As Djohan (2021) points out there is also an uncertainty in the socio-hydrological model itself and poor validation of the model at the farm level, which might also contribute to the lack of correlation. Another possible explanation is that there are other risk factors that haven’t been included in the risk matrix and that have a stronger impact on the crop yield. This indicates a need to further investigate the exact correlation between the currently chosen risk factors and the crop yield, as well as other potential risk factors. In order to draw any definite conclusions about the relationship between the variables, more farmers also need to be examined.

4.1.1. Sensitivity Analysis

In this section, the sensitivity analysis conducted for the shift in the monsoon onset and the change of the number of dry spells is discussed. However, the reader should first note that only two farmers

(29)

were examined for this part, and the results could have differed if more had been examined. If a more extensive sensitivity analysis were to be done it could be included as a feedback loop in the workflow process, presented in Figure 2.1, meaning that it could be used to assess the selection and definition of risk factors more thoroughly. Here, it still gives a suggestion of further work to be done, but does not directly affect the selection or definition.

The results from the sensitivity analysis for the shift in monsoon onset showed varying results on the yield. For the farmer in Mowada, the yield increased when the monsoon started later than usual, while it decreased for the farmer in Pandhari. A possible explanation for this might be that the farmer in Mowada experienced rainfall the days before the monsoon onset, which was not enough to trigger the detection of the start of the monsoon season, but was enough to increase the yield when these days later were moved to the start of the monsoon season. For the farmer in Pandhari, the rainfall might have been scarce before the onset of the monsoon, which could result in a lower yield when these days were moved to be the new onset.

The results from the sensitivity analysis for the dry spells showed no difference in yield when dry spells were added to the monsoon season of 2016 for the two farmers in Mowada and Pandhari. This could be explained by Figure 3.3 and 3.4 which display that the actual crop evapotranspiration for the years with added dry spells overlapped with the actual crop evapotranspiration for the original year.

Note that the actual crop evapotranspiration differed from the potential crop evapotranspiration in Figure 3.4 during some parts of the monsoon. However, since the deviation was the same for the years with added dry spells there is no reason to believe that this would affect the yield compared to the original year. This may indicate a possible weakness to the methodology in that the condition of 20 mm/week was not strict enough, meaning the crop would still have enough water available to not impede it’s growth. The method also only examined one week dry spells and not prolonged durations, which might have an equal or greater impact than the reduction of precipitation. Furthermore, the results added to the belief presented in the methodology that using an agricultural dry spell definition might be a better estimate of dry spell risk than the chosen metrological definition. As Barron et al.

(2003) argued, farmers may be more concerned with crop water stress than the number of consecutive dry days, since the impact on crop growth will depend more on agro-ecological conditions. This, together with the found result, suggests that there is reason to interview farmers in the chosen districts on their risk perceptions and not rely solely on literature, or to redefine and use an agricultural dry spell definition instead.

4.2. Risk Communication

Three options of communicating hydro-climatic risk to farmers in Maharashtra were proposed in the results section. The differences between the alternatives were not in the content but in the orientation of the figure and of the legend. Thus far, attention has been directed towards risk factors and probabilities, but it is equally important to emphasise the consequence, i.e. crop yield. Option 1.0, in which the figure is presented vertically, may be preferred if the user associates a higher and lower yield with a sense of “up” and “down”. Placing the plant icon on top of the bar chart may also provide symbolic meaning in that plants grow from the soil and up. Moreover, since both figure and legend are vertically oriented and placed next to each other, the user will be able to read the information

(30)

“top-down” and be assured that the first shade in the figure corresponds to the first key in the legend.

The choice of risk visualisation is also somewhat constrained by the layout of the user interface and of the size of the mobile screen. Since mobile devices are usually held standing, this would also favor option 1.0. Option 2, which had the figure placed horizontally, had two contenders. Both have the advantage of “readability”, compared to option 1.0, since people are used to reading from left to right.

The information also takes up less space than the vertical option, which might make the figure seem less “overwhelming”. Option 2.2. has the legend placed horizontally, which offers the same advantage as for option 1.0. However, the text would have to be quite small in order to fit on one row, which favors option 2.1., with a vertical legend, instead. All things considered, it is proposed that the bar chart is presented vertically, i.e. option 1.0. Ultimately, however, the front-end team at TU Delft can make the decision and any changes, as they see fit.

Moreover, the literature reviewed in this study has indicated that effective risk communication is as much bound by its users as it is by the distribution of data. While there is ample evidence of visualising risk with averages and standard deviations, few studies seem to have considered the representation of non normal distributions. This study hopes to have contributed with a method for translating such a distribution into a more recognisable format, that can be understood by people other than academics and statisticians. However, it is possible that the proposed risk visualisation still requires prior knowledge of risk factors, likelihoods and consequences. Regarding the verbal probabilities, there is also a possibility for translation issues. Some words may not be translatable into Marathi, which is the official language in Maharashtra, and even so, the words may not bear the same meaning. Concerns about the usability are of course magnified considering that the target users also may lack in literacy and/or numeracy. Although this study has assessed the needs of the audience and proposed ways to overcome these difficulties, it is ultimately unknown how the risk visualisation will be perceived and understood by the farmers in Maharashtra. For this reason, the objective of this study has not been entirely fulfilled. It can only be fully met once the design prototype has been tested with the farmers and iterated towards a final design.