A comparison of two methods to predict the landscape-scale variation of crop yield

Landscape-scale variation is a source of information that increasingly is being taken into consideration in agricultural and environmental studies. Models that encompass and interpret this variation in fields and across contrasting management practices have the potential to improve the landscape management of agroecosystems. Our objective was to compare the results of two approaches, analysis of covariance (ANCOVA) and state-space modeling, to determine the factors affecting grain yield in three crop rotations pea (Pisum sativum L.)–wheat (Triticum aestivum L.)–barley (Hordeum vulgare L.), canola (Brassica napus L.)–wheat–barley, and wheat–wheat–barley] at two sites in Saskatchewan, Canada. Crop rotations were established in adjacent 30 m×80 m plots arranged in a randomized complete block with five replicates. Variables that were expected to affect resource availability and pest infestations in wheat (second rotation phase) or barley (third rotation phase) were measured. Each sampling point was classified according to landscape position as either a shoulder or footslope. Landscape position was considered as a cross-classified treatment along with crop rotation, and analyzed using ANCOVA procedures. State-space modeling was conducted on a single transect connecting sampling points across all of the rotations and replicates at each site. ANCOVA frequently indicated that grain yield and other measured variables differed between landscape position across all rotations, or in a specific crop rotation. For example, grain yield, soil water content, soil N availability during the growing season, and the incidence of common root rot were higher in the footslopes than the shoulders in all of the crop rotations at one of the sites. However, the landscape position effect for grain yield was never fully explained by the landscape position effects detected for the other variables (e.g., higher soil water content in the footslopes did not correspond with higher grain yields in footslope positions at both sites). State-space modeling indicated that most of the measured variables contributed to the prediction of landscape-scale variation for grain yield in the pea–wheat rotation; whereas only leaf and root disease incidences explained landscape-scale variation in the wheat–wheat rotation. The selective omission of data indicated that state-space modeling was accounting for the varied importance of the predictors across the landscape; i.e., localized response functions. The major reason that ANCOVA did not explain landscape-scale variation of grain yield across the different crop rotations may be because it was unable to account for highly localized variation. However, there is evidence from other studies that the ANCOVA approach is appropriate when the response functions explaining grain yield do not vary significantly within the study area. This situation is most likely to occur in studies with smaller experimental areas. Future research conducted at scales reflecting ‘real world’ field conditions (i.e., study units representative of producer’s fields) should consider the use of state-space modeling or alternative statistical techniques that are designed to address and predict the complex and dynamic nature of landscape-scale processes.