Just out todayAugmented reality: The AR trainees were faster on the model they trained on, and no better anywhere elseVirtual reality: They matched the exercise dose exactly, and the VR advantage fell below the line that mattersVirtual reality: Both groups used the same headset. One watched a video first and picked up 10.8 more blocks

Agritech/Thailand

The strongest predictor of a Thai farmer’s rice yield was who they talk to

Given 58 variables from 1,722 Thai rice farmers, the model found that social network strength predicted yield better than any agronomic input.

What the paper found

A survey of 1,722 smallholder rice farmers in central Thailand with 58 agronomic and economic variables found that the most influential predictors of yield were farmers' social networks, rental costs during harvest and total harvesting expenses. AutoGluon achieved RMSE 0.532 tonnes per hectare and R-squared 0.538. Clustering on SHAP values revealed six distinct farmer cohorts.

Grade 12 reading level5 min read

What happened

Precision agriculture is built on physical inputs: soil chemistry, rainfall, fertiliser rates, satellite vegetation indices. Models are trained on those and recommendations follow from them.

A team at the Department of Agronomy at Kasetsart University in Bangkok, the Faculty of Economics at Chiang Mai University, and King Mongkut’s Institute of Technology collected something broader. They surveyed 1,722 smallholder rice farmers in central Thailand across 58 agronomic and economic variables, trained four automated machine learning systems, and then asked the model which variables mattered.

The strongest predictor was the farmer’s social network.

The test

Four automated machine learning frameworks were evaluated with five-fold cross-validation: AutoGluon, auto-sklearn, h2o and mljar. Using several rather than picking one is a reasonable guard against the result being an artefact of a particular algorithm.

AutoGluon performed best and became the basis for the analysis.

Importance was then assessed two ways. Global SHAP analysis identified which variables mattered most across all farmers. Clustering on individual SHAP values grouped farmers by what drove their particular yields – a more interesting question than overall importance, because it asks whether all farmers are limited by the same thing.

The numbers

Measure Result
Farmers surveyed 1,722 smallholders, central Thailand
Variables collected 58 agronomic and economic
Frameworks compared AutoGluon, auto-sklearn, h2o, mljar
Best model AutoGluon
RMSE 0.532 tonnes per hectare
MAE 0.372 tonnes per hectare
R-squared 0.538
Top predictors social networks; harvest rental costs; total harvesting expenses
Farmer cohorts identified 6
Data Zenodo, with DOI

The result

AutoGluon achieved root mean square error of 0.532 tonnes per hectare, mean absolute error of 0.372 and R-squared of 0.538.

Global SHAP analysis identified farmers’ social networks, rental costs during harvest and total harvesting expenses as the most influential predictors of rice yields. Stronger social network connectivity was associated with higher yields, which the authors interpret as information sharing and collective knowledge within farming communities improving productivity.

Clustering on individual SHAP values revealed six distinct farmer cohorts, each characterised by unique patterns of feature importance.

The authors conclude that combining automated machine learning with explainability techniques can move beyond uniform agricultural recommendations towards precision support tailored to different farmer cohorts.

How this compares to what came before

Yield modelling for smallholder systems usually relies on remote sensing and weather, because those data are available at scale and farm-level surveys are expensive. Models built that way can only find what they are given, and they are given physical variables.

By collecting 58 variables including economic and social ones from 1,722 farms, this study gave the model a chance to find something else, and it did.

The social capital finding is not new to agricultural economics – the role of farmer networks in technology adoption and knowledge diffusion is a long-standing research area with its own literature. What is new is that a machine learning model given a free choice across agronomic and economic variables ranked it first.

What it means

The headline implication is for where agricultural development effort goes. If social connectivity is the strongest predictor of yield among these farmers, then farmer field schools, cooperatives and extension networks are not soft complements to technical intervention. They may be the intervention.

That is a claim about association, and it needs care. Better-connected farmers plausibly differ in wealth, education, land quality and risk tolerance, and the survey may not capture all of those. The direction of causation is also not established: higher-yielding farmers may be more sought after as contacts.

The cohort finding is the more immediately usable result. Six groups with different yield drivers means uniform extension advice is necessarily wrong for most recipients. Segmenting advice by what actually limits each group is a concrete programme design change.

The modest R-squared should temper any use of this as a prediction tool. Explaining half the variation in smallholder yields is reasonable for survey data and not good enough to forecast a harvest.

Business ideas from this paper

Farmer network infrastructure. If connectivity predicts yield, then the product is connection – structured peer groups, knowledge-sharing platforms, cooperative formation – and that is cheaper to deliver than inputs.

Cohort-based advisory services. Six distinct groups with different drivers means six different advisory packages. Segmenting an extension service this way is operationally straightforward and currently nobody does it.

Harvest cost financing. Two of the three top predictors were harvest rental and total harvesting expenses. That points at a financing product timed to the harvest window rather than planting.

Survey instruments for agricultural ML. Most smallholder datasets collect agronomic variables only. Designing and running surveys that include economic and social variables, at scale, is the input that made this finding possible.

What could be wrong with it

R-squared of 0.538 means roughly half the variation in yield is unexplained. As a predictive model this is weak.

SHAP importance measures association within the model, not causation. Social connectivity may be a proxy for something else the survey did not record.

The data are self-reported by farmers in a single collection. Social network strength in particular is difficult to measure and depends entirely on how the question was asked.

Central Thailand is one region with one crop, one set of land tenure arrangements and one extension system.

No intervention was tested. The cohorts are identified and nobody has shown that targeting advice to them changes yields.

How sure can you be?

We score this 8 out of 10.

It is a peer-reviewed paper in PLOS ONE, based on a substantial farm-level survey of 1,722 smallholders across 58 variables, with three named institutions in Thailand – Kasetsart University, Chiang Mai University and King Mongkut’s Institute of Technology – and the dataset is openly deposited on Zenodo with a DOI.

It loses points because there is no independent replication and the evidence is cross-sectional survey analysis rather than an experiment.

The open data are a real strength. Anyone who disagrees with the modelling can download the survey and redo it.

Do this today

If you run an agricultural extension programme, look at how your advice is segmented. This study found six groups limited by different things, and the default is to tell everyone the same.

If you build models on farm survey data, include economic and social variables. A model given only agronomic inputs cannot find what this one found.

If you are reading yield prediction claims, check the R-squared before the variable importance. A model explaining half the variation can still rank its inputs informatively, and it cannot forecast your harvest.

Source: Explainable machine learning reveals diverse yield-determining factors among Thai rice farmer cohorts: Implications for targeted agricultural support, PLOS ONE, June 2026. PMID 42296117 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.

What to remember
  • 1,722 smallholder farmers surveyed across 58 agronomic and economic variables.
  • The top predictors were social networks, harvest rental costs and total harvesting expenses - not soil or inputs.
  • R-squared 0.538 means the model explains about half the variation, so this is a weak predictor and a strong finding about what matters.
  • Six distinct farmer cohorts emerged from clustering on individual SHAP values, each with different drivers.
Where this really is

Comprehensive farm-level data were collected from 1,722 smallholder farmers in central Thailand across 58 variables, and the dataset is openly deposited.

Questions people ask

What is SHAP?

A method that attributes a model's prediction to each input variable, so you can see which features drove which prediction. Global SHAP identifies overall importance; clustering on individual SHAP values groups farmers by what drives their yields.

Does social connectivity cause higher yields?

The study shows association, not causation. Better-connected farmers may differ in other ways the survey did not capture - wealth, education, land quality - and the authors interpret it as information sharing within farming communities.

Why does R-squared of 0.538 matter?

It means the model explains roughly half the variation in yield. As a prediction tool that is weak. As an exploration of what drives yield, the ranking of variables is still informative.

What are the six cohorts?

Groups of farmers whose yields are driven by different factors, identified by clustering on each farmer's individual SHAP values. The point is that one set of recommendations will not suit all of them.

About the author

Mark Alex

Mark Alex is the founder and Managing Director of Real Biz Digital, a technology company operating out of Nairobi since 2018. He works in agentic AI and the Model Context Protocol, AI governance, enterprise software architecture and cybersecurity. He holds an MSc in Mechatronical Engineering from Obuda University in Budapest and a BSc in IT, Forensic Technology and Cybercrime, from USIU-Africa in Nairobi, and has published IEEE conference research on an AI-powered digital twin for greenhouse systems. He is the author of seven books. Between 2020 and 2024 he mentored more than 200 university students and interns in Nairobi. He writes every Just Out Tech article from the original research paper.