Talk · IEEE International Conference on Smart Information Systems and Technologies · May 13, 2026

Multimodal Data Fusion for District-Scale Crop Yield Prediction in Northern Kazakhstan

Assylmurat Yeraliyev, Nurzhamal Kashkimbayeva, Dinara Kaibassova

Abstract

This paper predicts the yield of spring wheat at the district level in Kazakhstan's Shortandy district (Akmola oblast) by merging meteorological reanalysis data, MODIS vegetation indices, and ISRIC SoilGrids soil data into one machine learning algorithm. The dataset has 21 crop years$(2004-2024)$that collected from open APIs: Open-Meteo for climate data, NASA ORNL DAAC for NDVI and EVI, ISRIC for soil data, and the Bureau of National Statistics of Kazakhstan for yield data. We tested eight regression algorithms, including Ridge, Lasso, Random Forest, XGBoost, LightGBM, and SVR. We trained them on data from 2004 to 2020 and tested them on data from 2021 to 2024. The best performance was from a Ridge model with strong L2 regularization$(\alpha=10): R^{2}=0.246$, RMSE$=0.264 \mathrm{t} / \text{ha}^{-1}$, and MAE$=0.205 \mathrm{t} / \text{ha}^{-1}$on the four years that were not used for training. Precipitation totals and the minimum temperature during the growing season were the most important features, which makes sense given the dry conditions in northern Kazakhstan. We think that every tree-based model had a lot of overfitting because the district only reports one yield value per year, which means there are only 21 unique target observations in the whole dataset. We also made a web app using Flask that lets users enter weather data and get yield predictions. All data are drawn from public sources, and collection scripts are provided for reproducibility.

Remote Sensing in AgricultureSmart Agriculture and AIClimate change impacts on agriculturecrop yield predictionspring wheatmultimodal fusion

Multimodal Data Fusion for District-Scale Crop Yield Prediction in Northern Kazakhstan

Abstract

More by Nurzhamal Kashkimbayeva