The problem of variability in soil nutrient analysis has been studied for years by a number of industry experts; unable to decipher and commercialize hyperspectral soil sensing. Many studies have taken years of testing to account for variability thathas a dramatic impacts on precision of recommendations. The main tradeoff we have identified is between accuracy and precision. Large quantities of raw data are required to result in precise action. Acquiring large data sets using traditional soil analysis is prohibitively expensive. But with hyperspectral imaging, the same costs can result in far more data, dramatically increasing efficiency and precise input recommendations. Big data is often associated with multiple data sets (weather/climate, chemical soil tests, satellite imaging, LIDAR, yield monitoring, etc.) to translate this information into actionable recommendations. Large computing systems, validation withdiverse cropping systems, and a lot of time is required to createvalid products for true precision agriculture. Reliable data is essential to manage performance and identify areas of improvement. This means that models are only as good as the data that is used to create them. Hyperspectral soil analysis provides a cost-effective method to obtain more data sets per acre for the same cost to the grower. In one study we acquired 81 data sets versus 3 data sets for the same cost (add in-text citation referring to appendix about the study or data set). More data isn’t the solution, actionable data is the solution!
Soil diagnostic is an area of expertise. Accurate prescription maps are essential for effective VRT fertilizer application (Sawyer, 1994; Ferguson et al., 1996). Grid soil sampling has most frequently been used to develop these prescription maps (Mueller et al., 2001). Past research has indicated several technical and economic limitations associated with this approach. There is a need to keep the number of samples to a minimum while still allowing a reasonable level of map quality. However, Gotway et al. (1996) found that the optimum grid density may depend on the coefficient of variation. In many cases, where the spatial distribution is rather complex, much finer grid densities than those currently used commercially are required to produce accurate prescription maps. Mueller et al. (2001) indicated that a common commercial grid sampling scale of 100 m was grossly inadequate and that sampling at greater intensities only modestly improved prediction accuracy that would not justify the increase in sampling cost. Their data suggest that the use of the field average fertility values at their research field was not substantially worse than grid sampling. Schloeder et al. (2001) demonstrated that spatial interpolation of grid sampled data with limited sample size (n = 46) was mostly inappropriate. For most of their data sets the inability to predict, could be attributed to either spatially independent data, limited data, sample spacing, extreme values, or erratic behavior. Whelan et al. (1996) reported that in fields with less than 100 samples only very simple geostatistical methods such as inverse distance are appropriate. Sample sizes of 100 to 500 are needed for geostatistical methods such as kriging. Kravchenko and Bullock, (1998) studied several interpolation techniques, such as ordinary kriging, lognormal kriging, and inverse distance weighting, and found the best geostatistical methods to use depended on unique spatial properties in each field and could not be predicted in advance. McBratney and Pringle, (1999) reported that grid sampling at 20 to 30 m is generally needed when applying site specific management at a resolution of 20 by 20 m. Mallarino and Wittry (1997) reported that cells larger than 0.8 ha in size usually did not represent nutrient levels appropriately.
Field level studies have shown that organic C, total N, and NO3-N have spatial dependence and variation (Cambardella et al., 1994). Using the ratio of nugget to total semi variance to classify spatial dependence, organic C, total N, and NO3-N were strongly spatially dependent. Other studies have concluded that N uptake and crop response to N varies spatially within fields (Malzer, 1996; Dampney and Goodlass, 1997). Welsh et al. (1999) reported significant yield increases where 30% more additional N was applied to historically higher yielding parts of the field. Kachanoski et al. (1996) showed that optimal levels of N fertilization have spatial variability. The maximum yield increase and the economic yield increase over the check yield with no N applied were both strongly correlated with spatially optimal economic return of N (r=0.70 to 0.88). Variable rate application technologies enable farmers to adjust N rates to reflect these variations.
Many researchers have shown that soil test levels of P and K vary considerably within fields as well. Various studies (Cahn, 1994; Cambardella et al., 1994; Mallarino, 1996; Nolin et al., 1996; Penney et al.,1996) have shown coefficients of variation (CV) of 30–55% for P and 19–43% for K. McGraw (1994) reported that of 392 fields sampled in western and southern Minnesota using grid sampling methods, the range of soil P and K values encompassed four to five soil-test interpretation classes in 86% (for P) or 61% (for K) of the fields. Furthermore, the spatial structure of soil-test variability often is site specific and nutrient specific (Mallarino, 1996; Borges and Mallarino, 1998).
Multiple studies in Montana proved the economics and accuracy of hyperspectral analysis to replace wet lab chemistry for soil testing. The first four year study resulted in a lime application to mitigate the impact of acidic soils. Lime was applied on 107 acres vs 220 as a result of more data sets to variable spread the lime. The savings generated by the use ofhyperspectral soil analysis was $21,879 in lime, $1,739.75 in lab savings, while receiving 69 sample points rather than 37 sample points. Cost savings is only part of the value, Crop Productivity Index or measure of variability with 100 being the mean value for that cropping season. The results for variability showed as:
2018 – 24 pt difference in CPI values
2021 – 4 pt difference in CPI values
Went from over 47% of the field with below optimal pH in 2016 to 1.6% of the field below optimal this season.
As can be seen, no one grid size or interpolation technique adequately describes the variability that exists in fields of a diverse population. If one fails to sample at a fine enough resolution to capture the spatial correlation in crop nutrient data, the interpolation methods and application maps developed from those methods will not be valid or accurate (Reich, 2000). However, the cost associated with grid sampling to the intensity required for accurate maps will be prohibitive in many caseswithout the use of hyperspectral soil analysis.
