Precision nitrogen management: site-specific management grids, soil nitrogen estimation, and optimum sampling

Abstract

Integrating macro-scale variability in soil and micro-variability in crops and translating it into a practical management unit is a complex task. Past studies on precision nitrogen (N) management have focused primarily on employing large-scale variability in soil. More recently, sensing N in crop canopies to characterize micro-scale variability has been used for customized fertilizer applications. However, macro- and micro-scale variability have not been effectively coupled and adapted to match current fertilizer application technologies (i.e., individual nozzle level control and application of fertilizer). Likewise, temporal and micro-spatial uncertainty of data for crop variables, influenced by environmental and crop management practices, have not been adequately addressed. A pre-requisite to most N prescription maps is quantification of soil NO₃-N spatial distribution in a field, which mandates intensive soil sampling and analysis. The complex mobility of nitrogen in soil, coupled with time, labor, and costs of analysis, underscores the necessity for an alternative method. Application of non-imaging hyperspectral sensing, characterized by high spectral resolution (1-3 nm bandwidth) of a spectroradiometer, facilitates a more comprehensive spectral analysis. This approach would have the potential to estimate surface soil NO₃-N rapidly and inexpensively without compromising accuracy in soil characterization. Most often, commercial soil sampling designs uses a simplified random and/or stratified design, which entails collection of soil samples without accounting for spatial dependency. In instances where ancillary data such as proximal soil and grain yield are available, leveraging a sampling technique that accounts for geographic and feature space optimization of sampling configuration would be essential. In this context, spatial representations emerge as a critical facet of site-specific management protocols, strategically designed to account for inherent complexities of within-field heterogeneity and spatial autocorrelation. The objectives of this study were to: (i) delineate site-specific management grids that account for macro-scale variability in soil and micro-scale variability in crop; (ii) estimate surface soil NO₃-N using non-imaging hyperspectral sensing; and (iii) optimize soil sampling configuration with conditioned Latin Hypercube Sampling (cLHS) technique and hybrid cLHS-Sparse Sensor Placement Optimization for Reconstruction (SSPOR) techniques. Studies were conducted for each objective across two to four locations in Kansas and Colorado spread over four to six site years. For objective 1, a hybrid fuzzy inference system (H-FIS) was developed to delineate SSMGs that account for macro-scale variability in soil and micro-scale variability in crops. A proximal soil sensor, Veris-MSP3 was used to acquire soil variables. The Unmanned Aerial Vehicle (UAV) equipped with on-board MicaSense Rededge-3 multispectral sensor was used to collect multi-spectral images during the crop growing season. Adaptive Neuro-Fuzzy Inference System (ANFIS) was used to generate a tuned membership function (MFs) within the H-FIS framework. Crafted fuzzy rules and expert knowledge combined fuzzy rules were defined with H-FIS model output to generate SSMGs. A combination of selected soil and crop variables and crop grain yield was used to tune MFs with R² > 0.7. The H-FIS model was used to couple micro-scale variability in soil and macro-scale variability in crop with the selected soil and crop variables (Elevation, Slope, ECa, OM, and Normalized Difference Red Edge (NDRE)). The H-FIS output data were used to map the SSMGs. A classification with ten classes of grid-based management units was successfully delineated for all site years. A similar spatial trend was observed between SSMGs delineated based on coupled macro-scale variability in soil and micro-scale variability in crops and grain yield (corn and soybean) for all site years. Overall, the complex coupling of soil and crop variables accounting for both macro- and micro-scale variability was achieved. Each management grid represented a distinct area with specific productivity potential characterized into multiple classes. These classes are expected to be used as a ‘weighted-coefficient’ based on SSMGs to adjust N prescription maps and optimize applications of the N rate. For objective 2, soil samples were analyzed for surface soil NO₃-N and spectral measurements were acquired with ASD FieldSpec3 spectroradiometer. Spectral preprocessing techniques were applied to refine the spectral data and remove noise. Effective spectral windows were determined such that spectral demonstrated sensitivity to changes in soil NO₃-N. The selected spectral windows: W1 (1460-1690 nm), W2 (1730-1780 nm), and W3 (1940-2150 nm) were used to estimate soil NO₃-N. Partial Least Squares Regression (PLSR) and Random Forest Regression (RFR) were used to estimate soil NO₃-N using selected spectral features and measured soil NO₃-N. Estimated surface soil NO₃-N maps were developed. Map of measured soil NO₃-N showed a similar spatial trend with estimated soil NO₃-N maps. The findings of potential hyperspectral data-based surface soil NO₃-N estimation models and maps are expected to provide a cost-effective and practical technique to improve soil fertility management practices. For objective 3, an optimized soil sampling design was developed by integrating multiple years of grain yield data with proximal soil data (with Veris-MSP3 soil sensor). The hybrid cLHS-SSPOR and cLHS techniques with proximal soil+yield data sources were used to optimize soil sampling sites. Results with different optimal soil sample sizes for site year-1 and site year-2 indicated the need for site-specific sampling strategies hybrid cLHS-SSPOR and cLHS techniques. In this study, Kullback-Leibler divergence of hybrid cLHS-SSPOR-based technique indicated its capability to characterize local spatial heterogeneity compared to the cLHS. These findings emphasize the importance of tailored sampling techniques for different study areas to accurately generate a soil sampling design that can characterize soil spatial variability. Overall, the results of this study introduced a data fusion method based on evidential reasoning that integrates macro-scale variability in soil and micro-scale variability in crops to delineate SSMGs. Non-imaging hyperspectral that uses spectroradiometer proves the potential of developing an accurate surface soil NO₃-N estimation model using selected spectral features. The importance of tailored sampling techniques was emphasized for different study locations to accurately generate a soil sampling design.