Spatial variation, bias and experimental design in agronomic field trials: a case study of a farming systems trial in the Western Cape province, South Africa
Thesis (MScAgric)--Stellenbosch University, 2017.
ENGLISH SUMMARY: Detecting small statistically significant changes in an observed parameter (e.g. – yield) in response to applied treatments within field trials is a major challenge for modern agriculture, since we continue looking for smaller increments where spatial variability (noise) may obscure the sought trend. A reliable detection of change can significantly reduce the length of some long-term trials required for a trend confirmation. This study assesses the effects of spatial variation in yield, terrain and soil properties within a long-term field trial on the possible outcome of results. The experiment was set up at the Langgewens research farm of the Western Cape Department of Agriculture, South Africa as an extension of the long-term trial initiated by the provincial government. In year 1 (2015) the whole field was planted to wheat as a uniformity trial to assess the magnitude of productivity, variation and spatial trends within the 12 ha experimental site. A yield monitor mounted on a combine harvester and remote sensing techniques using drone surveys were used to study the variability in wheat growth/yield, and topographic parameters within the field trial area. Soil profiles positioned along the field perimeter were interpolated into a soil map to understand some of the reasons for yield variability. The relationship between topographic parameters and wheat yield was significantly more important in the drier year (2015) than in 2016. Results from a random forest regression model indicated that the use of spectral vegetation indices NDVI and SAVI, in combination with topographic parameters can be used to accurately predict wheat yield in dry and in wet years (R2 = 0.974 and R2= 0.987 respectively). Wheat growth and productivity during the uniformity trial showed high variation within the trial area (1.46±0.61 t·ha-1) with a bimodal distribution largely determined by variation in topography and soil types. The first step in trial layout optimization was the exclusion of a small low-yielding area on the eastern margin retaining the bimodal distribution, but slightly increasing the mean and significantly reducing the standard deviation (1.68±0.46 t·ha-1). The reduced-area experimental setup comprises 120 plots combining three main 10-year crop rotation treatments starting at a different year of rotation resulting in k=30 x n=4 trial for each year and finally combining into k=3 x n=40 at the end of the trial period. Soil core samples at a 0-5cm depth (120 in total) were analysed for some key soil parameters. The variation in routinely analysed soil properties across the field within the 0-5cm layer, where the most changes may be expected as a result of applied treatments collectively showed no effect on wheat productivity. The only exception was the resistance measurements, which indicated a high degree of leaching negatively correlated with wheat production. The trial was further optimized for means by complete randomization of the trial setup by placing three treatment replicates in the high-yielding block and the fourth replicate in the lower-yielding part of the field. This resulted in a relatively uniform distribution of means and standard deviations across the trial. Such setup also allowed the possibility of replication reduction to n=3 (or n=30 at the end of the trial) by excluding the lower-yielding section. The supposition that reduction of replications under the circumstances may lower the difference was supported by power analysis. The power analysis of the 2015 yield has shown that three replications within a more uniform field allow for detection of smaller changes (160 kg·ha-1) as opposed to highly variable conditions with 4 replications (314 kg·ha-1). A different approach – optimization for within-treatment standard deviation was realized by quantifying and accounting for the spatial trend. The combination of replications into ranked groups results in significant reduction of within-group standard deviation and, subsequently smaller detectable differences. This approach should be further investigated in the time-series analysis as the current experiment progresses.