Abstract

Data-driven approaches hold great promise for better understanding of within-field crop yield variability. Data-driven models use spatiotemporal information that relate to plant growth and can be correlated with yield. Such data is available at different spatial resolutions so a decision is needed about what spatial resolution we should model at which can be determined based on the required management resolution; however, this does not always match the resolution of the input data. Prediction quality can change with spatial resolution. Therefore, this work focused on exploring changes in prediction quality with changes in the spatial resolution of predictors and the predictions. More specifically, this study investigated whether inputs should be resampled prior to modeling or the modeling implemented first with aggregation of predictions happening as a final step. The study consisted of 59 fields over six years in the south-west region of Australia. Rainfed canola, barley and wheat were used for this study. Data used for modeling were acquired from different sources. These data consisted of interpolated yield monitor data and soil sensor data, Landsat 8 data, elevation data, digital soil maps, and weather data. The first approach was based on resampling the inputs to four spatial resolutions (10, 30, 90 m and whole-field-average). The second approach was based on aggregating the outputs from the models which were built at 10, 30 and 90 m to the coarser resolutions and eventually to the whole-field-average. This study showed that the second approach (modeling at the highest resolution and aggregating the outputs to the whole field average) resulted in the highest Lin's Concordance Correlation Coefficient (LCCC) (0.87) and lowest RMSE (0.59 t ha(-1)). This study confirmed that better results can be obtained from machine learning models by following the appropriate approach to related to the spatial resolution resampling.