Machine Learning Algorithms for Spatial Interpolation of Crop Yield at the Field Scale

Abstract

Abstract Crop mapping is one of the most frequent uses of yield monitor data in precision agriculture. Processing such data requires spatial interpolation to predict yields at unsampled locations. Ordinary kriging (OK) is the geostatistical interpolation method usually employed for mapping georeferenced data. In recent years, machine learning (ML) algorithms have gained attention for spatial interpolation tasks because of their ability to process large data volumes; however, their comparative performance for yield mapping at the field scale remains limited. This study compares the statistical behavior of ML methods, including quantile regression forest (QRF), generalized boosted regression models (GBM), extreme gradient boosting (XGB), and radial basis function neural networks (RBFNs) for processing yield monitor data. OK is applied as reference. To enhance the performance of ML algorithms for fine-scale yield mapping, covariates from the spatial neighborhood of sampled yield values were incorporated to predict yields at unsampled sites. Over 1,000 yield monitor datasets from multiple crop species were processed to assess algorithm performance. The ML methods—QRF, GBM, and XGB—demonstrated robust statistical performance, since they effectively handled large data volumes and improved spatial interpolation accuracy. The QRF method achieved the highest error reduction (on average, an 8.7% error reduction), was faster than OK, and generated high-quality uncertainty maps. GBM and XGB also performed better than OK. Coupled with spatial covariables, the studied ML algorithms are valuable alternatives to conventional kriging for yield mapping at the field scale.

Plant phenotyping relevance

作物収量という植物形質を対象に、収量モニターデータの空間補間手法を比較・評価し、未観測地点の収量推定性能を検証しているため、計算的フェノタイピング手法が中心である。

Code and data availability

The supplied blocks describe ML spatial interpolation of 1,050 yield monitor datasets from the Argentine Pampas, but contain no data availability statement, no public repository deposit of the yield datasets, and no author code/workflow URL. All URLs in the text are bibliographic references to prior work or generic R/C

No evidence-backed public reproduction asset is currently recorded.