AI monsoon forecasts: Predicting India's rainy season for 38 million farmers

The problem

For millions of smallholder farmers across the tropics, when to plant is a high-stakes decision, shaped above all by the timing of the rains. Plant too early, and a dry spell can kill the seedlings. Plant too late, and the growing season may not be long enough for the crop to mature.

In India, the summer monsoon, which typically begins in southern India in June and moves north through July, determines whether most farmers have a good year or a bad one. Monsoon climates affect the lives of nearly two-thirds of the world's population, making this one of the most widespread climate-dependent decisions on earth. Much of India's agricultural workforce consists of smallholder farmers, who are amongst the most economically vulnerable communities in the world and "climate change is really threatening their livelihoods", as University of Chicago economist Michael Kremer has noted. The 2025 season illustrated the growing challenge when the monsoon arrived early in southern India, leading many to expect an early season everywhere, but then it stalled for nearly three weeks before moving again.

The problem is that farmers have had to make this decision largely without reliable advance information. While the India Meteorological Department issues a national forecast for when the monsoon reaches Kerala, at the southern tip of the country, research has shown this does not reliably predict when the rains will arrive in other parts of the country.

Without local forecasts, farmers are left to watch for the national monsoon declaration and observe conditions around them. "I mostly relied on my own experience and local knowledge to know when the monsoon would arrive," said Parasnath Tiwari, a farmer from Madhya Pradesh. This means that millions of farmers making the highest-stakes decision of the year were primarily relying on experience and observation.

The solution

Local predictions that give farmers weeks of advance warning rather than days had not been available. To close that gap, the Indian Ministry of Agriculture and Farmers' Welfare ran a programme built on the judgement that a 30-day forecast of the monsoon's arrival would help farmers decide when to plant. To build it, the Ministry partnered with an international team of researchers led by the University of Chicago's Human-Centered Weather Forecasts Initiative, working alongside the Indian Institute of Technology Bombay, the Indian Institute of Science Bangalore, and the University of California, Berkeley. As Pramod Kumar Meherda, Additional Secretary at the Ministry, put it, "This program harnesses the revolution in AI-based weather forecasting to predict the arrival of continuous rains, empowering farmers to plan agricultural activities with greater confidence and manage risks".

The team tested seven AI weather models against nearly 60 monsoon seasons of data, to determine which combination was most accurate at predicting local monsoon onset with meaningful advance warning.

Two models stood out. The first was Google Research's Neural GCM, an open-source model that integrates established atmospheric physics with neural network techniques to simulate weather patterns. The second was the Artificial Intelligence Forecasting System developed by the European Centre for Medium-Range Weather Forecasts. Both outperformed other AI and conventional models at predicting monsoon onset.

The researchers did not use either model on its own. When predicting two to four weeks ahead, the AI models alone struggled to add meaningful information beyond what a farmer could already observe on the ground. So the team built a blended model that combines AI weather predictions with a statistical component they designed, called the evolving-expectations model. This component is trained on 124 years of historical rainfall data from the India Meteorological Department, the national weather agency, and estimates when a farmer can expect rain based on how far into the season it has progressed without rain.

This component accounts for something the AI models on their own did not: as the season progresses without rain, a farmer already knows the monsoon has not arrived, and their expectation of when it will come naturally shifts forward. By building that real-world knowledge into the forecast, the blended model ensured that what farmers received was genuinely new information beyond what they could already see for themselves. This blended model, not any single component, was the forecast delivered to farmers.

The Indian Ministry of Agriculture and Farmers' Welfare deployed the forecasts to farmers via SMS through its existing messaging platform. Message design and testing were led by Precision Development, a global nonprofit that worked directly with farmers to determine the form the information should take.

One practical feature of the AI models used is that they do not require supercomputers to run. As University of Chicago researcher Mayank Gupta noted, "They can be run on desktops and can be tuned to the specific weather conditions and needs of the citizens on the ground, all at a fraction of the cost and time".

The impact

1. 38 million farmers received locally tailored monsoon forecasts up to 30 days in advance

The Indian Ministry of Agriculture and Farmers' Welfare delivered the forecasts to 38 million farmers by SMS across 13 states. The Odisha state government reached nearly one million additional farmers through a voice messaging platform. The forecasts were not broad national announcements. They told farmers in each area when sustained rains were likely to reach them, with probability estimates rather than single dates, so that farmers with different circumstances could make their own decisions about when and what to plant.

2. The forecast accurately predicted an unusual monsoon pause

The 2025 season was a particularly demanding test. Early rains in the south suggested a fast-arriving monsoon, and many farmers across the country began preparing accordingly. But in late May, the monsoon's northward advance stalled for around 20 days, an unusual pause that left farmers who had already started planting exposed to a prolonged dry spell.

The blended model accurately predicted this. Weeks before the pause happened, the forecast warned farmers that the rains would stall. Standard forecasting methods, which work by comparing the current season to historical averages, failed to anticipate it. When a season breaks from past patterns, those averages mislead. The blended model held its accuracy throughout. For farmers, advance warning of a dry spell after initial rains matters: planting during what appears to be the start of the monsoon, only for the rains to stop, can destroy a crop.

3. Farmers changed their decisions based on the forecasts

For farmers like Parasnath Tiwari in Madhya Pradesh, the forecast changed how they planned. Tiwari used the advance warning to get a head start on the season and opted for higher-earning crop varieties that he would not have risked without the forecast's reassurance about season length. He also passed the forecasts on to other farmers in his area, sharing how he has "increased trust in the forecast, and [I] will rely on the information shared by scientists in the future". Previous research by the University of Chicago across 250 villages in Telangana found that when farmers received accurate monsoon-onset forecasts a month in advance, they meaningfully changed their behaviour. Some reduced farming activity in unfavourable seasons and started non-agricultural businesses, nearly doubling their business profits in the process.

4. The blended model outperformed every individual model tested

The blended model outperformed all individual AI models and all standard multi-model averages across every evaluation metric and time period tested. According to the research paper, it achieved a roughly 15% improvement in accuracy at one-week lead times compared to the evolving expectations baseline alone, and remained more accurate than climatology out to four weeks. The model's accuracy was validated across 60 years of historical monsoon data and held up during the abnormal 2025 season.

Lessons learned

• A model that performs well scientifically may still not be worth sending. The research team established a clear principle before the programme launched: forecasts should only be delivered if they contain actually new information beyond what farmers already know. At longer lead times, the AI models alone did not clear that bar. Scientific accuracy and practical usefulness are not the same thing, and the gap between them matters most for the people the service is designed to reach.

• Farmer input shaped what the technology needed to do. Before any models were tested, the team worked with farmers to understand what decisions the forecast needed to support and what form the information should take. Farmers wanted probability estimates, not single dates. They wanted enough lead time to act. The technology was selected and designed to meet those requirements.

• Existing government infrastructure was sufficient for delivery. The programme reached tens of millions of farmers without building anything new. For governments in other countries considering similar programmes, that is a significant finding: the barrier to delivery may be lower than it appears if existing communication channels are already in place.

• The underlying models are publicly available and do not require specialist infrastructure. Both AI models used in the blended forecast are open-source and can run on a desktop computer. The research team described this as an opportunity for technological leapfrogging, allowing countries that lack expensive computing infrastructure to still access forecasts.

Launch date: 2022

This case study was written with assistance from artificial intelligence.