The use of yield prediction maps is an important tool for the delineation of within-field management zones. Vegetation indices based on crop reflectance are of potential use in the attainment of this objective. There are different types of vegetation indices based on crop reflectance, the most commonly used of which is the NDVI (normalized difference vegetation index). NDVI values are reported to have good correlation with several vegetation parameters including the ability to predict yield. The field research was conducted in two commercial farms of processing tomato crop, Cantillana and Enviciados. An NDVI prediction map developed through ordinary kriging technique was used for guided sampling of processing tomato yield. Yield was studied and related with NDVI, and finally a prediction map of crop yield for the entire plot was generated using two geostatistical methodologies (ordinary and regression kriging). Finally, a comparison was made between the yield obtained at validation points and the yield values according to the prediction maps. The most precise yield maps were obtained with the regression kriging methodology with RRMSE values of 14% and 17% in Cantillana and Enviciados, respectively, using the NDVI as predictor. The coefficient of correlation between NDVI and yield was correlated in the point samples taken in the two locations, with values of 0.71 and 0.67 in Cantillana and Enviciados, respectively. The results suggest that the use of a massive sampling parameter such as NDVI is a good indicator of the distribution of within-field yield variation.

The processing tomato is one of the most important crops in Spain, producing around 1.97 million tonnes. In recent years, the management regime of this crop has undergone a series of changes as a result of an increase in average field size. New tools are consequently required to enable a global view of these larger-sized fields and to determine the heterogeneous zones that often appear within them. The use of yield prediction maps is an important tool for the delineation of within-field management zones. In particular, appropriate placement in the field of organically grown produce, as is the case of the processing tomato crop in this study, will result in higher crop yield. Yield prediction maps are of great importance to ensure that the crop is harvested at the right time and that production yields are maximized for industrial processing (

The field research was conducted in two farms, “Los Enviciados” (-7.009427 38.950592 decimal degrees) and “Cantillana de Mesas” (-6.942781 38.946368 decimal degrees), with study areas of 6.50 ha and 7 ha, respectively. The farms are situated in the proximity of Badajoz (southwest Spain). The climate of this area is characterized by variation in both temperature and precipitation typical of a Mediterranean climate, with mean annual precipitation of less than 500 mm. One of the most important characteristics of the precipitation is its interannual variability. There is a dry season, from June to September, and a wet season, from October to May (80% of the precipitation falls between these months). Summers are hot, with temperatures sometimes rising above 40ºC.

The field was transplanted with processing tomato (

The NDVI survey was conducted in August 2013, 12 days before harvesting, with a Crop Circle ACS-470 reflectance sensor (Holland Scientific Inc., Lincoln, NE, USA) held by a tractor at a height of 80 cm above the canopy. The Crop Circle ACS-470 generated reflectance data in the wavebands 670 (red) and 760 (NIR) which were combined to obtain the NDVI following equation:

is a reflectance value of the waveband (760 nm) and

An ARVATEC monitor with Topcom GB500 GPS and JAVAD GDD base with sub-meter accuracy was used to geo-reference the NDVI measurements. NDVI data at 10 second intervals were recorded on an ACS-470 data logger in an ASCII text format. Later, this raw ASCII file was transferred to other software for further analysis. NDVI measurements were made along different parallel transects approximately 8 m apart (yellow dots with the appearance of lines in

Firstly, an NDVI prediction map for each entire plot was developed using ordinary kriging technique [see

An exploratory analysis in which the data were studied without considering their geographical distribution and statistics were applied to check data consistency, removing any outliers and identifying the statistical distribution of the data.

A structural data analysis was developed, in which spatial distribution was evaluated using variograms of the variable (NDVI); the equation of the mathematical model, nugget effect (micro-scale variation or measurement error), sill (variance of the random field) and range (distance at which data are no longer auto-correlated) were used to develop these variograms.

The prediction maps were developed using a data search by neighbourhood copied from the variogram; the NDVI prediction maps were used to guide yield sampling.

A study was then made of the areas from which the samples were taken to determine the relationship between NDVI and yield properties (

where the trend, m(x), is fitted using linear regression analysis and the residuals, r(x), are estimated using ordinary kriging algorithm. If cj are the coefficients of the estimated trend model, vj(x) is the jth predictor at location x, p is the number of predictors and wi(x) are the weights determined by solving the ordinary kriging system of the regression residuals, r(xi), for the n sample points, then the prediction is made by:

In this case study, only one predictor is used, NDVI, so m(x) = a + b•(NDVI(x)). In consequence:

The residual at each sampling point, r (xi), is calculated as the difference between the value of the parameter and the estimate by the trend (r (xi) = Z (xi) – m (xi)).

The geostatistical analyses were conducted using the Geostatistical and Spatial Analyst extensions of the GIS software ArcGIS (v. 10.0, ESRI Inc., Redlands, CA, USA). All maps were produced with the ArcMap module of ArcGIS.

Finally, a comparison was made between the yield obtained at the validation points and the yield values according to the prediction maps (ordinary kriging and regression kriging). In order to assess the quality of the maps in terms of yield prediction we adopted the statistical procedure proposed by

where n is the number of observations, Pi is the predicted value, Oi is the measured value, and Õ is the mean of the measured values. Validation is considered to be excellent when the RRMSE is <10%, good if it is between 10% and 20%, acceptable if the RRMSE is between 20% and 30%, and poor if it is >30% (

The exploratory analysis of the data distribution in this work could indicate a normal distribution of the data, as it is indicated in the results chapter, corroborating the choice of kriging for this work. Kriging has been recommended as the best method to interpolate point data since it minimizes error variance using a weighted linear combination of the data (

There are no studies showing the benefits of geostatistical analysis techniques to develop NDVI-based yield prediction maps for the processing tomato, though numerous studies have shown the benefits of geostatistical analysis techniques for agricultural management (