INTEGRATING REMOTELY SENSED IMAGES WITH A SOYBEAN MODEL TO IMPROVE SPATIAL YIELD SIMULATION

Abstract

The CROPGRO-Soybean model has been used in the past to explain causes of spatial yield variability and to estimateoptimum prescriptions in soybean fields. However, the model requires many spatial inputs, which can be very expensiveto collect. Recently, researchers have used calibrated remotely sensed vegetative indices as a real-time adjustment to leafarea state variables to improve simulated yield variability by crop models. However, calibrated images are not readily availablefor commercial agriculture. Most commercial image providers distribute uncalibrated vegetative indices as a relativeindicator of plant health. The purpose of this work was to determine if within-season adjustment of model leaf weight statevariables using an uncalibrated normalized remotely sensed vegetation index improves simulation of spatial soybean yields.The CROPGRO-Soybean model was calibrated to simulate several years of historical spatial yield data in 98 cells 0.2 hain size for the Heck Home field, 77 cells 0.2 ha in size for the Heck McGarvey field, and 207 cells 0.07 ha in size for the Bakerfield in central Iowa. The model was then used to simulate spatial yield distribution for an independent season in each field.Initially, cell-level simulated yields generally did not follow the trend in measured yields, with r2 values ranging from 0.04to 0.47 for the best-fit line between simulated and observed yield for the testing year. A procedure was developed to adjustsimulated leaf weight based on a normalized vegetation index (NVI) from an uncalibrated image in each cell. Spatial soybeanyields were simulated in the independent season again using this adjustment in leaf weight based on the NVI in each cell. Boththe slope of the best-fit regression line for simulated versus measured yield and the correlation coefficient (r2) improved inall three cases. The slopes of the best-fit line for simulated and observed yields for the Heck farm fields were not statisticallydifferent from one, and the intercepts were not statistically different from zero. The NVI correction did not improve theaccuracy of simulated yields for the independent season in the Baker field. This was likely due to the image being taken duringpod filling rather than near flowering, which was too late for model feedback from the NVI adjustment to have an effect onpod addition. These findings indicate that uncalibrated vegetative indices can improve simulated yield variability in somecases. More research is needed to determine the optimum timing of NDVI measurements leading to the most accuratesimulated yield variability.