There is increasing demand for more detailed soil maps to support fine-scale land use planning, soil carbon management, and precision agriculture in Saskatchewan. Predictive soil mapping that incorporates a combination of environmental covariates provides a cost-effective tool for generating finer resolution soil maps. This study focused on mapping soil properties for multiple soil horizons in Saskatchewan using historical legacy soil data in combination with remote sensing band indices, bare soil composite imagery, climate data, and terrain attributes. Mapped soil properties included soil organic carbon content (SOC), total nitrogen, cation exchange capacity (CEC), electrical conductivity (EC), inorganic carbon (IOC), sand and clay content, and total profile soil organic carbon stocks. For each of these soil properties, a recursive feature elimination was undertaken to reduce the number of features in the overall model. This process involved iteratively removing features such that random forest out-of-bag error was minimized. Final random forest models were built for each property and evaluated using an independent test dataset. Overall, predictive models were successful for SOC (R2 = 0.71), total nitrogen (R2 = 0.65), CEC (R2 = 0.46), sand content (R2 = 0.44) and clay content (R2 = 0.55). The methods used in this study enable mapping of a greater geographic region of Saskatchewan compared to those previously established that relied solely on bare soil composite imagery.
A three-dimensional predictive soil mapping approach for predicting soil organic carbon (SOC) stocks (t/ha) at high spatial resolution (30 m) for Alberta for 2020–2021 is described. A remote sensing data stack was first prepared covering Alberta’s agricultural lands. A total of 404 sampling locations were distributed across Alberta using 2-scale sampling: (1) 22 pilot farms representing main climatic zones and (2) conditioned Latin hypercube sampling at each farm. Soil samples were taken at four standard depths (0–15, 15–30, 30–60, 60–100 cm) using soil probes and analyzed for SOC. Predictive models for SOC content and bulk density were built separately and then used to predict at 0, 15, 30, 60, and 100 cm and calculate aggregated SOC stocks per pixel. The SOC content and bulk density models had R squares of 0.61 and 0.68, respectively. Based on these mapping results, grassland soils were consistently associated with higher SOC stocks across all soil types as compared to croplands. The average SOC stock increase for grassland soils compared to cropland soils was 2.1 Mg per hectare, ranging from 2.17 to 6.09 Mg per hectare depending on soil type. Results also showed that >15 % of total SOC stocks were located in subsoil, which was higher than expected.
26 December 2023 in Soil Science Society of America Journal
Utilizing reflectance spectroscopy to generate the necessary soil data to drive innovations in precision agriculture and soil management is an increasing focus of agronomic research. One of the key limitations for widespread practical adoption of reflectance spectroscopy is hardware cost, and lower cost hardware is actively being developed. This study evaluated two inexpensive nano Fourier‐transform near infrared spectrometers in the laboratory. One was a laboratory‐based analyzer (LabFlow) and the second was a field portable analyzer (Field Probe). Soil spectra were collected in the shortwave infrared range and processed using wavelet transforms and machine learning models. The optimal wavelet transforms and machine learning model were selected using cross validation on the training dataset, and performance of the optimal model was evaluated using an independent testing dataset. The Field Probe configuration total nitrogen model had the best performance when compared to the LabFlow laboratory analyzer with an R2 of 0.91, a concordance correlation coefficient of 0.95, and an root mean square error of 0.03. Soil inorganic carbon did not perform as well with an R2 of 0.65. However, performance was likely limited by a large number of low values and a limited range in the training dataset. Overall, these results highlight the potential for lower cost spectrometers to provide useful soil data for soil management applications.
The major drivers of soil variation in Saskatchewan at scales finer than the existing soil maps are parent material variance, slope position, and salinity. There is therefore a need to generate finer-scale parent material maps as part of updating soil maps in Saskatchewan. As spatially referenced soil point data are lacking in Saskatchewan, predictive soil mapping methods that disaggregate existing soil parent material maps are required. This study focused on investigating important environmental covariates to use in parent material disaggregation, particularly bare soil composite imagery (BSCI). Synthetic point observations were generated using an area-proportional approach based on existing soil survey polygons and a random forest model was trained with those synthetic observations to predict parent material classes. Including BSCI as environmental covariates increased model accuracy from 0.38 to 0.52 and the model Kappa score from 0.19 to 0.35 compared with models where it was not included. Models that included training points from all locations, regardless of whether BSCI was available, and included BSCI as environmental covariates had similar results to the BSCI model with an accuracy of 0.48 and a Kappa value of 0.30. Based on these results, BSCI is an important covariate for parent material disaggregation in the Saskatchewan Prairies. Future work to disaggregate soil classes based on slope position and salinity, and to combine those methods with parent material disaggregation is needed to generate detailed soil maps for the Canadian Prairies.
Despite the increasing use of portable, low-cost spectrometers in estimating soil properties, there is lack of documentation regarding the factors contributing to the lower performance of these spectrometers when compared to conventional ones. This study investigates potential factors influencing performance of the Nanoquest, a low-cost spectrometer, in estimating soil organic carbon (SOC) and total nitrogen (TN). To conduct the study, five different models (cubist, partial least squares regression, support vector machines, random forest, and generalised boosted models) were tested for the estimation SOC and TN and a fivefold cross-validation analysis was conducted for model hyperparameter optimization. The Nanoquest achieved a Lin’s concordance correlation coefficient (CCC) value of 0.84 and an R2 value of 0.74 for SOC. For TN, CCC values of 0.86 and an R2 value of 0.78 were obtained. To understand the impact of the spectral range and spectral resolution on SOC and TN estimation, the ASD spectra were digitally resampled to match the Nanoquest spectral range and resolution. This resampling resulted in a slight decrease in model performance for the spectral range and a more pronounced decrease for the spectral resolution.
Wetland soil types, which can be distinguished based on calcium carbonate content, vary in their effect on ecosystem functions like phosphorus retention, salinity contributions, and greenhouse gas forcing. These soil types may be predictively mapped with machine learning models that use terrain derivatives calculated from high-resolution digital elevation models. Soil profiles from three Saskatchewan study sites were classified into three functional categories—upland, calcareous wetland, or noncalcareous wetland—and used to train random forest models for predictive soil mapping. Multiple terrain derivatives were included as predictor variables to capture local- and landscape-scale morphometry and hydrology influences, including five derivatives developed for this study. Models were developed at three spatial resolutions: 2, 5, and 10 m, and tested via internal cross-validation and independent validation with datasets from previous studies. Predictive accuracies were highest when mapping at 2 m resolution (independent validation accuracy range = 64%–100%) but also successful when mapping at 5 and 10 m resolutions (independent validation accuracy range = 63%–100%); however, visual inspection determined that the maps generated at 10 m resolution were less detailed and occasionally featured questionable discontinuous soil distributions. Three of the five terrain derivatives developed for this study were among the most important predictor variables (first, second, and 10th most important). Models trained using only data from a specific site had slightly better performance than models trained using data from all sites, except in regions where training data were lacking.
About Me
Dr. Sorenson is a research associate at the University of Saskatchewan.
Dr. Sorenson is a research associate at the University of Saskatchewan.