Greig Farm Soil Carbon Testing Summer 2021

In early summer 2021, Skidmore College students took soil samples from 153 locations over 120.4 acres at the Greig Farm in Red Hook, NY, at an average sampling density of 1 sample per .79 acres. At each location, they took two samples: one sample from 0-15 cm below the surface and one sample from 15-30 cm below the surface. Measuring soil carbon to a 30 cm depth is IPCC standard. Ward Laboratory in Kearney, Nebraska, analyzed the samples for percent carbon by dry combustion. This method does not distinguish between organic and inorganic matter and is considered one of the more accurate methods of testing.

Based on the following data, we estimate that ~2,560 tonnes of carbon were contained in the soils of the sampled fields at Greig Farm.

We analyzed results on average and at each depth per sample location. Where the students were unable to sample due to ground cover, moisture, or compaction, results were omitted.

-Mean percent carbon in soil 0-30 cm below surface was 1.39%. Mean percent carbon in soil 0-30 cm below surface in non-plowed areas was 1.55%.

-Mean percent carbon in soil 0-15 cm below surface, 1.66%, was higher than mean percent carbon in soil 15-30 cm below surface, 1.17%. This decrease is consistent with common soil patterns.

-Differences in percent carbon with ground cover and plowing were observed in soil at 0-30 cm and 0-15 cm but not in soil at 15-30 cm. Highest percent carbon in soil 0-30 cm in non-plowed areas approached 3.5%. Highest percent carbon values in soil at 0-30 cm in plowed areas approached 2%. Because percent carbon was higher in soil at 0-15 cm than in soil at 15-30 cm, patterns in soil at 0-15 cm likely significantly contributed to mean values across soil at 0-30 cm.

We did not sample soil bulk density. Estimates of total carbon by mass require density estimates. Bulk density can be sampled at lower intensities than percent carbon. Some methods use regression models to estimate bulk density by applying a commonly observed correlation between decrease in bulk density with increase in percent carbon. We suggest sampling to confirm models before application. Decreases in bulk density may produce an accrual of ecosystem co-benefits, including increased soil porosity, water infiltration, and microbial activity.

Summary

Data returned in October 2021. Across Greig Farm’s sampled fields mean percent carbon at 0-30 cm was 1.39% with a standard deviation of .44% (Table 1). Mean percent carbon at 0-15 cm was 1.66% with a standard deviation of .59%. Mean percent carbon at 15-30 cm was 1.17% with a standard deviation of .46%.

Data was non-normally distributed, per Shapiro-Wilks tests, with positive skews across the soil strata (Fig 1). This distribution determined subsequent statistical tests.

Spatial variation

We fit semivariogram models to percent carbon by sample location to observe baseline local spatial variability and distance over which samples appear to be related. Semivariance is a mathematical characterization of variation between all points a specific distance apart. Results show that percent carbon values spatially relate to each other across soil samples taken at 0-15 cm and 0-30 cm (Fig. 2). Percent carbon values do not significantly spatially relate across soil samples taken at 15-30 cm. At that soil depth, a pattern may be present but it is not statistically significant across all samples. Baseline semivariance in the fitted semivariogram of percent carbon at 15-30 cm suggests that there is higher local percent carbon variability at this depth than on average and at 0-15 cm. 

To further characterize spatial autocorrelation or relationships, we calculated the Moran’s I statistic across soil depths. This statistic represents similarity per sample proximity across all samples. If samples close to each other are similar, this test returns a statistically significant result.The Moran’s I statistics for these values support a somewhat small amount of spatial autocorrelation in the top layer of soil. Percent carbon values from soil at 0-15 cm had a Moran’s I statistic of .0675 with a p-value of 0. Percent carbon values from soil at 0-30 cm had a Moran’s I statistic of .0215 with a p-value of .0028. The Moran’s I test was not significant for samples from soil at 15-30 cm, suggesting that there is not significant spatial autocorrelation across those samples. These results support the fitted semivariograms above.

A different pattern of percent carbon distribution in soil at 15-30 cm may be the product of underlying soil type. Research suggests soil type significantly drives soil carbon retention in lower soil horizons, which is often more mineral, or inorganic. It may also be related to historical land use.

Estimate of total carbon

We combined the mean value of percent carbon sampled at a high intensity on Greig Farm with USDA bulk density estimates per sampled area soil type to estimate total carbon contained in Greig Farm’s sampled fields. 

USDA’s Web Soil Survey summarizes the measured fields per mapped soil type (Fig. 3, Table 2). We note that these soil types are mapped to a scale of 1:24,000, and are not reliable on smaller scales, including this one. Bulk density values per soil type do not take land use into account.

We estimate that the sampled fields at Greig Farm contain approximately 2,490 tonnes of carbon. The EPA’s Greenhouse Gas Equivalencies calculator estimates this carbon as the equivalent of 1,986 passenger vehicles driven for one year, or 1,027,343 gallons of gasoline consumed.

Variation across land use

We ran Kruskal-Wallis tests with Bonferroni corrections to test whether percent carbon varied with ground cover and plowing. Ground cover was reported in field by the Skidmore students. We ran Kruskal-Wallis tests based on the data’s non-normal distributions. There were significant differences in percent carbon per ground cover and plowing.

In soil at 0-30 cm, areas under tall grass, grass, orchard, and blueberries had statistically higher percent carbon than areas under blackberries and strawberries (Fig. 4). Areas under tall grass also had statistically higher percent carbon than areas under no cover, asparagus, and cover crops. In soil at 0-15 cm, areas under tall grass, blueberries, orchard, and grass had statistically higher percent carbon than areas under no cover, strawberries, and asparagus. Areas under tall grass also had statistically higher percent carbon than areas under blackberries and cover crops. There was no statistical difference in percent carbon across areas with different ground cover in soil at 15-30 cm.

Areas managed without plowing had statistically higher percent carbon than areas managed with plowing at 0-30 cm (Fig. 5). Mean percent carbon in non-plowed areas was 25% higher than that of plowed areas (Table 3). Mean percent carbon in non-plowed areas was 46% higher than mean percent carbon in plowed areas at 0-15 cm. Evaluating soil carbon variation as a difference in means is not sufficient. There is overlap in the spread of the data from both management areas. Mean values of percent carbon from the non-plowed areas are, in part, pulled higher by samples of particularly high percent carbon. Highest percent carbon from areas managed without plowing at 0-30 cm was nearly 3.5%. Highest percent carbon from areas managed with plowing at 0-30 cm were approximately 2%.

There was no statistical difference in percent carbon between areas managed with and without plowing at 15-30 cm.

Lastly, we ran a stepwise linear regression of percent carbon from 0-15 cm by SSURGGO soil structural characteristics and land use. We used SSURGGO values of percents sand, silt, and clay, weighted average per soil type. We tested what would produce the best regression model of percent carbon distribution between combinations of percents sand, silt, and clay, management with plowing, and ground cover/crops. The analysis dropped the soil structural characteristic variables and management with plowing to produce a simple linear regression that predicted percent carbon by ground cover. R2= .36, F(10, 136) = 7.7, p<0.0001. We checked model fit by the residuals and qqplot, having log-transformed the percent carbon values to produce a normal distribution.

We suspect that the dropping of soil structural characteristics reflects small scale inutility of SSURGGO soil type information. As the SSURGGO products state, the mapped soils are not considered reliable at small scale. Two soil types comprised most of the sampled fields.

We note that percent sand and percent silt do have significant but weak relationships with percent carbon values. As shown previously via Kruskal-Wallis tests, as does management with plowing. Since the stepwise regression excluded these variables from its model of best fit, we infer that the information they contributed was not novel: it was comprised in the ground cover. This is confirmed by strong relationships between these variables.

Between SSURGGO-reported percent sand, silt, and clay, management with plowing, and ground cover, ground cover alone is the best predictor of percent carbon in a given location. We do notsuggest that ground cover is the driver of this percent carbon: this data cannot address that. However, ground cover not only reflects the larger variations recorded by SSURGGO, but also smaller, field-scale differences.Here we infer: ground cover will reflect underlying site conditions, based on farmers’ choices, and may also change those conditions.

There is a danger in overfitting or overestimating the fit of models by using multiple variables that share the same underlying drivers. We note too that percent silt and percent sand on this property were very nearly inverse of each other, and almost one variable repeated. Variation in SSURGGO-reported percent clay was minimal and thus the variation was likely too small for effective statistical analysis.

To accurately describe an impact of management, including plowing, there needs to be data from an analogous site or system, in a different state. To quantify impacts of differences in management with plowing due to plowing on Greig Farm or similar, one would want to compare two areas of the same soil and ground cover, managed with and without plowing.

Maps of percent carbon results on Greig Farm

ArcGIS Online interpolated maps of percent carbon on Greig Farm

ArcGIS Online-interpolated maps visualize patterns from the statistical analysis above. We recommend their use as visual aids but not as models (we are planning to move out of ArcGIS Online). Interpolations of percent carbon at both 0-30 cm and 0-15 cm, below, show higher interpolated values of percent carbon in areas managed without plowing than in areas managed with plowing.

ArcGIS Online maps of mean squared prediction error

ArcGIS generates maps of prediction error when it runs interpolations. These values represent statistical error, which is a mathematical value and not the degree to which something is “wrong” in the vernacular sense. Mean squared error is the average square of the difference between data and a predicted value (the interpolated value) at a given point. Therefore, visualizations of mean squared prediction error show, in part, local spatial variability. They are in the same units as the measured variable. This error value generally increases as percent carbon increases.

We reiterate that the interpolations above should be used as visual aids, not predictive models.

Thank you for welcoming us to Greig Farm.