Precision agriculture for South Asia

by Judith Curry

An exciting new project for my company, Climate Forecast Applications Network (CFAN) to support smallholder farmers in Pakistan and India.

I have been working on a lot of interesting projects for clients of my company.  I haven’t written about them on the blog since nearly all of the projects are proprietary (for companies) and for the others I simply don’t have time to write them up.  This particular project is a public one, and our clients have written 3 very nice blog posts on the project.

Our client is Precision Development (PxD). PxD is a global non-profit organization that builds low-cost information systems at scale to share knowledge with the world’s poorest and most disadvantaged people.  I am extremely impressed by what PxD is doing and their personnel, and we welcome this project as a way to return to CFAN’s roots – our first project was flood forecasting in Bangladesh [link]

Apart from the technological challenges of calibrating weather forecasts with very little observational data, there are some very interesting challenges in working with the PxD team to understand how farmers perceive weather forecasts and how to make forecasts more useful to them.

The idea of providing high quality weather forecast information to smallhold farmers who are living on the edge is important for development and for adaptation to climate variability and change.   Operational adaptation, which plans in advance to manage in the face of adverse weather conditions, can be important tool for helping these farmers build some modicum of wealth, and avoid the endless impoverishment of having their crops destroyed by adverse weather conditions.  Through this project, we are helping to develop this capability.

Below are extended summaries of three blog posts from PxD:

Weathering the storm: Smallholder farmers and the untapped potential of weather forecasts [link]

Accurate weather forecasts would reduce uncertainty for farmers, yet they are under-supplied and under-studied in developing countries. Government and market failures reduce the quality and reach of weather information in rural regions, reducing its value for households. To address these gaps, Precision Development (PxD) is piloting the provision of improved forecasts to farmers in India and Pakistan. The pilot will investigate how farmers interpret and use weather information, including the cognitive and informational barriers that constrain adoption.

Smallholder farmers live with copious amounts of risk, leaving them vulnerable to income variability and losses. Weather uncertainty is a major source of this risk. A survey of farmers in poor, southern districts of India, for example, found that 73% of respondents had abandoned their crop at least once, in the ten years prior to being surveyed, after misjudging the onset of the monsoon; and one quarter had replanted.

Because weather variability can be disastrous for poor households and is increasing with climate change, economists have dedicated much attention to products that mitigate its consequences, including index insurance and climate-resilient crop varieties.  While effective in increasing farmers’ investments and profits, many of these innovations incur high distribution costs that limit their scalability. PxD is therefore exploring an alternative solution for farmers that face increasing climate risk: accurate, phone-based weather forecasts.

Forecasts have a high potential for cost-effectiveness and scale. By reducing uncertainty over future conditions, they allow farmers to make better production decisions throughout the season. Long-range forecasts of conditions over one to three months can enable farmers to make more informed decisions about how much to invest or which crops to grow, while shorter-range products can be used to determine the best time to conduct activities like fertilizer application. These decisions have the potential to generate meaningful yield impacts; PxD’s agronomy team in India estimates that transplanting rice seedlings from the nursery to the field at the right time can generate yield increases of up to 10%, relative to transplanting too late. Moreover, the marginal cost of delivering phone-based forecasts is low relative to other products that reduce risk, meaning that the returns to their provision would increase with scale.

Currently, however, rigorous evidence on weather forecasts’ effects on farmer outcomes is sparse. Part of the reason may be the limited availability of accurate, useful forecasts in many developing countries. The limited evidence base suggests that these quality shortfalls reduce forecasts’ value for farmers. Short lead times and a lack of complimentary advice may further explain the lack of impact.

PxD’s new pilot aims to build the evidence on the benefits of improved forecasts for smallholders. We have partnered with the Climate Forecast Applications Network (CFAN), a private provider with expertise in developing innovative weather information tools in South Asia, to deliver hyper-local forecasts and related advisory to smallholders in India and Pakistan. CFAN’s products are anticipated to improve on existing forecasts in the region in terms of accuracy, lead times, and precision. Importantly, they can also be calibrated to predict weather phenomena that are particularly relevant for agricultural decision-making, such as prolonged dry spells suitable for fertilizer application or monsoon onset dates.

PxD’s pilot activities will test the efficacy of different intervention designs in both contexts, laying the groundwork for a large-scale randomized evaluation of weather forecasts’ effects on agricultural outcomes that we hope to implement in 2023. Some of the questions we will explore during this preliminary phase include:

  • To what extent do farmers update their subjective weather expectations in response to different forecasts?  What role do informational, cognitive, trust or other barriers play?
  • How do farmers interpret probabilistic weather information? Do forecasts with a numeric probability of a weather event and a qualitative likelihood of the same event affect their beliefs differently?
  • For which agricultural decisions could improved forecasts generate the greatest returns?
  • Does improved weather information spread among farmers within and across villages?

The findings of the pilot and subsequent evaluation will have policy implications at PxD and beyond. For example, improved understanding of the returns to information products that reduce agricultural risk will inform PxD’s prioritization of new services, such as pest prediction models. The project will also seek to provide policymakers with evidence and incentives to increase funding to improve forecasts in regions that are under-served by governments and markets.

Brewing better weather services for Indian coffee farmers [link]

Coffee is a notoriously fickle crop. For example, heavy rainfall can damage crops, result in premature fruit-drop, increase the incidence of pests, and wash away fertilizer with negative implications for plant nutrient levels. Increasing weather variability and the incidence of extreme weather events associated with climate change will have significant negative effects on coffee producers. Given the sensitivity of the crop — and yields — to fluctuations in the weather, coffee farmers are likely to derive meaningful benefits from accurate and timely weather forecasts.

Insights from PxD’s Coffee Krishi Taranga (CKT), digital advisory service for small coffee farmers in India in partnerxhip with the Coffee Board of India, will be used to inform the design of a larger evaluation of the weather-integrated service and to scale an enhanced service to over 150,000 coffee farmers across four Indian states (Karnataka, Kerala, Tamil Nadu, and Andhra Pradesh).

More accurate information about medium-term rainfall — with a lead time of up to 15 days — will enable farmers to make informed decisions about applying nitrogen fertilizer and increase the likelihood that they apply this input during dry spells to reduce run-off and leaching. Similarly, if farmers are alerted to impending heavy rain, they can leverage this information to alter harvesting times or take other precautionary measures to protect crops and insulate yields.

In interviews conducted with coffee growers in August 2021, only 16% of respondents reported accessing forecasts. Integrating weather information into CKT’s existing services will broadcast weather forecasts to farmers tailored to their specific contexts and complement these forecasts with agronomist-designed advice.

Studies conducted in other contexts find that farmers form subjective expectations about upcoming weather events based on various factors, including their past experiences, local rules of thumb, existing forecast information, the costs and benefits of acquiring such information, and perceptions about how relevant weather-related risk is to their incomes. These expectations inform behavior over the course of the coffee crop cycle as farmers make decisions relating to input and investment choices, the timing of activities, and so on. The sum of these decisions, in turn, influences outcomes that farmers (as well as researchers and practitioners) are interested in – notably plant health, yields, costs, and profits. The goal of this research is to understand each of these elements through measurement and service pilots, A/B tests, qualitative interviews, and in-person workshops with farmers, agronomists, and extension agents.

The objective of the first set of interviews is to better understand how coffee farmers make decisions relating to the timing of agronomist-identified, weather-dependent coffee activities: fertilizer and lime application, coffee pruning, shade regulation, and harvesting. We hope to identify how weather fits into these decisions and what other factors influence the timing of these activities. If other limiting factors (such as the availability of an input) impact timing to a greater extent than the weather, forecast information with short lead times may not help farmers optimally time their practices without access to complementary inputs or information. These interviews will also help us identify how farmers interpret weather forecasts they already have access to, what impact incorrect forecasts have on their activities and on their trust in forecasts and the extent to which farmers discuss their expectations of upcoming weather with other members of their communities.

Coffee farmers in a subset of villages in our three study districts will then be invited to participate in in-person workshops, where they will interact with different forecasting formats. The workshop will be in the form of a ‘lab-in-the-field experiment’, where participants engage with an interactive platform that presents weather forecasts together with incentivized agricultural decision-making scenarios. Utilizing participants’ decisions on the platform, an ‘in-scenario’ weather ‘realization’ will be simulated, allowing participants to accrue a higher payoff for a ‘better’ decision. The best-performing forecast will accrue the highest cumulative payoff across participants and will inform our understanding of which forecast formats most effectively aid decision-making. The ‘best-performing’ customized-to-context weather forecast will then be piloted in the field among a sample of existing CKT users to evaluate whether it improves decision-making in a real-world setting.

Cottoning on: A free weather product for Punjab’s cotton belt [link]

PxD is building a free weather forecast product for farmers in Pakistan’s Punjab Province. The service will ultimately serve 490,000 smallholder farmers across the province.

To guide more informed product design decisions, in November the Pakistan team commenced with a set of end-user interviews with 55 cotton and wheat farmers. While 71% of wheat farmers cited access to weather forecasting information, only 45% of cotton farmers surveyed reported access to weather information. When asked if weather information “helped in planning”, 88 and 86% of cotton and wheat respondents respectively responded in the affirmative. Farmer users were asked to “Please list the types of weather challenges you have experienced” in the three years prior to being surveyed. Forty-three percent of respondents who reported experiencing weather-related challenges cited heavy rainfall and 30% reported high winds.

These types of weather incidents can be very costly for smallholder farmers with limited resources. Inundation washes away expensive inputs, creates mud that blocks sprouting crops, and creates conditions conducive to disease, while wind can destroy or severely damage crops throughout the cropping cycle. Further, when prompted to answer “Out of these weather challenges, which three have the largest impact on your crop costs and yields”, pest management was reported as the most common answer. Many pests and diseases thrive in particular weather conditions and can proliferate quickly if conditions are optimal. This suggests that a combination of advisory and weather forecasting information alerting farmers to be on the lookout for pests and to initiate pesticide application decisions sooner could be valuable as a means for reducing pest damage.

Another core challenge in developing this product is the quality of the weather forecasts themselves. The existing forecast services in Punjab tend not to be designed with the end user’s needs in mind. Three services in Punjab, Pakistan exemplify usability issues: the first requires user-initiated, user-paid inbound calls; the second requires internet access; and the third is only available to subscribers of a particular telecoms company.

Learnings from these research activities, coupled with insights generated by our collaboration with CFAN, will inform the final design of our Kharif weather forecasting product for cotton farmers. Once this product is up and running, the focus will shift to impact evaluation, including pick-up rates, behavior changes, and forecast accuracy. Pending a review of such impact outcomes, and funding considerations, the product will be further developed to cover the Rabi season and potentially scaled to additional countries.

This Kharif season, we are thrilled at the prospect of hundreds of thousands of Punjabi smallholder cotton farmers witnessing less fertilizer – and by implication, fewer resources – washed away or harvests inundated by unexpected rains, deploying more effective irrigation tailored to the forecast, and incurring fewer crop losses due to heavy rains and winds. When we asked our PxD Pakistan agronomist about the potential impact of this product, she said: “This product will reduce farming risks and expenses, and increase what is usually a farmer’s sole source of income – allowing for more investment into farming or money for personal expenses”. 

CFAN’s precipitation forecasts

CFAN is providing a comprehensive range of forecast variables, including temperature, wind, soil moisture, humidity, cloudiness, and thunderstorms.  However, the single most important forecast variable is precipitation.  CFAN is using 15 day ensemble forecasts from both ECMWF and NOAA (GEFS).

CFAN’ provides probabilistic weather forecasts based on the forecast ensemble (e.g. 51 ensemble members from ECMWF for each forecast).  The ensemble forecasts are transformed into meaningful probabilities through calibration of the forecasts by eliminating biases and distributional errors in the ensemble forecast (see schematic below)

Screen Shot 2022-06-26 at 10.39.59 AM

Forecast calibration requires good observational data.  Particularly in Pakistan, there is very little useful data from surface observations.  For rainfall, we elected to use the satellite-based IMERG precipitation data, which does the best with extreme rainfall events.

While the ECMWF forecast system does quite well, our calibrations significantly increase forecast skill.   CFAN provides a deterministic ‘best guess’ forecast  Below are two forecasts for single locations in the Punjab

Screen Shot 2022-06-26 at 11.13.29 AM

CFAN’s probabilistic rainfall forecasts are provided in terms of probability of exceedence of certain threshold daily rainfall amounts: 0.25 mm, 2 mm, 5 mm, 10 mm and 20 mm.  Below is shown the value add from CFAN’s calibration for the threshold 0.25 mm, which is basically the rain-no rain threshold.   CFAN’s calibration (red) provides substantial improvement to the skill relative to the native ECMWF forecasts (black), especially for shorter lead times.

Screen Shot 2022-06-27 at 7.36.06 AM

The skill score used is the ROC, which is explained below

Screen Shot 2022-06-27 at 7.36.46 AM


Seasonal forecasts

Our clients have also requested that we provide seasonal forecasts of monsoon rainfall.  These forecasts help guide decisions on which crop varietals to plant, and how much to plant.  CFAN is using a statistical-dynamical approach to provide a probabilistic seasonal forecast.  While our approach shows skill, sometimes our forecast will be wrong.  Our clients have asked how they can effectively make use of our seasonal forecasts, given that sometimes the forecast will be wrong.

Consider the following example from the chapter on Decision Making Under Deep Uncertainty in my forthcoming book:

Consider the following decision that farmers face annually. The farmer has a choice between two crops: Crop #1 produces a steady yield under all rainfall conditions, while Crop #2 provides large yield only under conditions of high seasonal rainfall. Since we can’t reliably predict rainfall on seasonal time scales, Crop #1 is the safest choice, although yield will be suboptimal if rainfall is high. Robustness becomes an important decision criterion when the consequences of making a wrong decision are high.  If crop insurance is available to protect against potentially poor yields, or if sufficient savings are available, Crop #2 may be the best strategy. If these tools and resources are not available and the consequences of a few years of low yields would be disastrous, then robustness becomes the priority. An alternative strategy is to hedge by planting mostly Crop #1, but devoting a fraction of the land to Crop #2.

Over the season, each farmer faces a range of weather-related decisions over different timescales.  While substantial confidence can be placed in CFAN’s rainfall forecast at 3 day lead times, confidence in the forecasts decreases with increasing lead time, with relatively low confidence in the seasonal forecast.  Hedging strategies are well matched to probabilistic weather forecasts.  We will be working with the PxD team to consider how hedging strategies could be implemented in response to the probabilistic weather/climate forecasts.

via Climate Etc.

June 27, 2022 at 09:49AM

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s