Establishing field water balances is difficult and costly, the variability of their components being the major problem to obtain reliable results. This component variability is presented herein for a coffee crop grown in the Southern Hemisphere, on a tropical soil with 10% slope. It was observed that: rainfall has to be measured with an appropriate number of replicates; irrigation can introduce great variability into calculations; evapotranspiration, calculated as a remainder of the water balance equation, has exceedingly high coefficients of variation; the soil water storage component is the major contributor in error propagation calculations to estimate evapotranspiration; and that runoff can be satisfactorily controlled on the 10% slope through crop management practices.
Sampling field soils to estimate soil water content and soil water storage (S) is difficult due to the spatial variability of these variables, which demands a large number of sampling points. Also, the methodology employed in most cases is invasive and destructive, so that sampling in the same positions at different times is impossible. However, neutron moderation, time domain reflectrometry, and, more recently, frequency domain reflectrometry methodologies allow measurements at the same points over long time intervals. This study evaluates a set of neutron probe data, collected at 15 positions placed randomly along a coffee crop contour line, over 2 years at 14-day intervals. The temporal stability of S was again demonstrated, so that wetter or dryer locations remain so over time, and the definition of such positions in the field reduces the number of sampling points in future S evaluations under similar conditions. An analysis was made to determine the minimum number of sampling points to obtain the average S of the field within a chosen level of significance. Classical statistical analysis indicated that the 15 measurement positions could be reduced to four or even to one position to obtain a reliable field S average. State–time analysis showed S estimations depend more on previous measurements of rainfall P (52%) than on evapotranspiration ET (28%) and S (20%). The analysis also showed that ET was not realistically estimated from previous measurements of S; it was more dependent on previous measurements of ET (59%) than on P (30%) and S (9%). This statistical procedure showed great advantages over classical multiple regressions. Future studies of this type should be carried out at regularly spaced observation points in a grid, in order to allow a 2-D and 3-D state–space–time analysis.