Tree Regression Analysis to Determine Effects of Soil Variability on Sugarcane Yields

Approximately 15% of south Florida sugarcane (Saccharum spp.) is grown on high water table sandy soils that overlie limestone bedrock. This study determined treatment and site-specific factors affecting sugarcane production on these soils using a new statistical tool called tree regression. Sugarcane was grown in a 38-ha area for three seasons (1991, 1992, and 1993). Treatments were subirrigation water table depth (0.45 vs. 0.70 m), N fertilization frequency (13 vs. 7 split applications for 3 yr at 224 kg N ha ' yr 1 ), and Mg fertilizer rate (0 vs. 60 kg Mg ha 1 yr -1 ), using a split-split plot design. Soil was sampled from plots before each crop to determine pH, and soil test P, K, Ca, Mg, and Si. Depth to rock was determined with ground-penetrating radar. Three statistical techniques were used to examine design and the effect of soil factors on sugarcane yield: traditional simple correlations, general linear mixed-model analysis (GLM and MIXED), and a new technique, tree regression. Tree regression resulted in functions encompassing the complexity of response between yield, soil nutrients, and other factors, while handling large amounts of data. The regression tree identified sugarcane yields ranging from 42.6 to 100.8 t ha 1 , grouped according to conditions defined by soil Ca, crop, soil Mg, the P intensity/capacity ratio, and water table level. The strength of the general linear mixed-model approach was in inference testing, whereas the strength of tree regression tree analysis is for prediction of covariate importance under broadly spaced environments.