Using Digital Agriculture Methodologies to Generate Spatial and Temporal Predictions of N Conservation, Management and Maize Yield
thesisposted on 03.01.2019 by Min Xu
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
The demand for customized farm management prescription is increasing in order to maximize crop yield and minimize environmental risks under a changing climate. One great challenge to modeling crop growth and production is spatial and temporal variability. The goal of this dissertation research is to use publicly available Landsat imagery, ground samples and historical yield data to establish methodologies to spatially quantify cover crop growth and in-season maize yield. First, an investigation was conducted into the feasibility of using satellite remote sensing and spatial interpolation with minimal ground samples to rapidly estimate season-specific cover crop biomass and N uptake in the small watershed of Lake Bloomington in Illinois. Results from this study demonstrated that remote sensing indices could capture the spatial pattern of cover crop growth as affected by various cover crop and cash crop management systems. Soil adjusted vegetation index (SAVI), enhanced vegetation index (EVI) and triangular vegetation index (TVI) were strongly correlated with cover crop biomass and N uptake for low and moderate biomass and N uptake ranges (0-3000 kg ha-1 and 0-100 kg N ha-1). The SAVI estimated cover crop biomass and N uptake were +/- 15% of observed value. Compared to commonly used spatial interpolation methods such as ordinary kriging (OK) and inverse distance weighting (IDW), using the SAVI method showed higher prediction R2 values than that of OK and IDW. An additional advantage for these remote sensing vegetation indices, especially in the context of diverse agronomic management practices, is their much lower labor requirements compared to the high density ground samples needed for a spatial interpolation analysis.
In the second study, a new approach using the multivariate spatial autoregressive (MSAR) model was developed at 10-m grid resolution to forecast maize yield using historical grain yield data collected at farmers’ fields in Central Indiana, publicly available Landsat imagery, top 30 cm soil organic matter and elevation, while accounting for yield spatial autocorrelation. Relative mean error (RME) and relative mean absolute error (RMAE) were used to quantify the model prediction accuracy at the field level and 10-m grid level, respectively. The MSAR model performed reasonably well (absolute RME < 15%) for field overall yield predictions in 32 out of 35 site-years on the calibration dataset with an average absolute RME of 6.6%. The average RMAE of the MSAR model predictions was 13.1%. It was found that the MSAR model could result in large estimation error under an extreme stressed environment such as the 2012 drought, especially when grain yield under these stressed conditions was not included in the model calibration step. In the validation dataset (n=82), the MSAR model showed good prediction accuracy overall (± 15% of actual yield in 56 site-years) in new fields when extreme stress was not present. The novel approach developed in this study demonstrated its ability to use elevation and soil information to interpret satellite observations accurately in a fine spatial scale.
Then we incorporated the MSAR approach into a process-based N transformation model to predict field-scale maize yield in Indiana. Our results showed that the linear agreement of predicted yield (using the N Model in the Mapwindow GIS + MMP Tools) to actual yield improved as the spatial aggregation scale became broader. The proposed MSAR model used early vegetative precipitation, top 30 cm soil organic matter and elevation to adjust the N Model yield prediction in 10-m grids. The MSAR adjusted yield predictions resulted in more cases (77%) that fell within 15% of actual yield compared to the N Model alone using the calibration dataset (n=35). However, if the 2012 data was not included in the MSAR parameter training step, the MSAR adjusted yield predictions for 2012 were not improved from the N Model prediction (average RME of 24.1%). When extrapolating the MSAR parameters developed from 7 fields to a dataset containing 82 site-years on 30 different fields in the same region, the improvement from the MSAR adjustment was not significant. The lack of improvement from the MSAR adjustment could be because the relationship used in the MSAR model was location specific. Additionally, the uncertainty of precipitation data could also affect the relationship.
Through the sequence of these studies, the potential utility of big data routinely collected at farmers’ fields and publicly available satellite imagery has been greatly improved for field-specific management tools and on-farm decision-making.