Consider two invented 90-day seasons, each divided into three equal windows. Their average daily maximum temperatures are 32, 26 and 26°C in one season, and 26, 26 and 32°C in the other. Both have a whole-season average of 28°C. A model given only that average cannot tell when the heat arrived. This arithmetic example assigns neither season a yield.
Splitting the growing season into three windows improved all four algorithms in a rice-modelling study published on 13 October. Beyond three, gains mostly flattened. The targets were CYGMA-simulated yields.[1]
The simulator keeps a crop calendar
CYGMA already represents time in its crop calculations. Its 2017 description advances growth day by day, measuring development against accumulated growing-degree days and the crop’s thermal requirements. It calculates how much light the canopy intercepts, converts that into biomass and allocates a share to the harvestable crop.[2]
The model also applies daily stress from heat, cold, nitrogen shortage, too little water or excess water. The strongest stress constrains that day’s growth. Management enters too: nitrogen inputs, irrigation and changing crop varieties affect what the same weather can produce. A weather-only learning model is therefore being asked to recover part of a richer process from compressed inputs.[2]
The researchers drew 100 training samples of 1,000 grid-years from 1987–2016 and tested against 10,149 observations from 2017–2019. Their full dataset contained 111,639 simulated grid-years. Seven weather variables were averaged over one to five equal-length seasonal segments.[1]
What the extra windows bought
Across those resolutions, random forest’s R² ranged from 0.37 to 0.56; XGBoost’s ranged from 0.20 to 0.44. Both outperformed linear regression and linear-kernel support-vector regression.[1]
R² compares squared prediction errors with the errors from using the test-set mean. A higher value means the predictions account for more of the variation in these simulated yields. It does not mean that a stated percentage of individual harvest predictions is correct.
Equal calendar windows only approximate crop development. The timing result comes from a simulator; transferring it to harvest forecasting requires observed-yield validation.[1]
Sources & further reading
- Effects of within-season temporal aggregation of weather variables on performance and interpretability in rice yield machine learning models Scientific Reports · 13 Oct 2026 · Accessed 19 Oct 2026
- Responses of crop yield growth to global temperature and socioeconomic changes Scientific Reports · 10 Aug 2017 · Accessed 19 Oct 2026