Optimization of geospatial data modelling for crop production by integrating proximal soil sensing and remote sensing data

Emerging technologies in precision agriculture (PA) offer a wide array of advanced methods to assess soil properties and to determine soil variability. Remote sensing (RS) and proximal soil sensing (PSS) technologies, widely used in quantifying surface and subsurface soil parameters, can be combined to infer spatial patterns of soil heterogeneity and to develop thematic maps for site-specific management. However, the use of these soil sensors must be reviewed constantly to maintain their efficiency and precision in delineating the soil-crop relationship and to inform PA approaches. Data mining and model optimization are key to evaluating high-density geospatial data in a dynamic production system. High-density PSS and RS-based soil characterization was explored and optimization techniques for digital soil mapping in PA were evaluated.In a first study, sensor measurements were subjected to multivariate statistical analysis, followed by an evaluation of a new Neighborhood Search Analyst (NSA) and the capacity of other data clustering algorithms to delineate spatially contiguous zones in agricultural fields and to optimize soil sampling locations to inform best management practices. PSS-based topography, apparent electrical conductivity (ECa), and RS-based indices data from 3 sites in Ontario, Canada, were employed to assess the novel technique’s performance in accurate zone delineation. In creating homogeneous zones, a maximum of 70% field variance (R2 = 0.70) was achieved. The R2 of the k-means cluster compared to that of the NSA was relatively higher (R2 = 0.80) where, the k-means cluster map consisted of groups or pixels with isolated boundaries in various parts of the field. The NSA’s unique capacity, across various locations, to produce an optimum (or user-defined) number of zones highlighted its superiority to k-means’ partitioning with isolated boundaries.A second study assessed the utility of PSS-based soil characterization in developing an optimum prediction method for multiple soil properties at 12 sites across Ontario, Canada. Measured ECa, topographic parameters and six lab-quantified soil properties [pH, buffer pH, soil organic matter (SOM), Phosphorus (P), Potassium (K) and Cation Exchange Capacity (CEC)] were used in evaluating the method’s predictive capacity and to compare different fields’ propagated soil measurement errors by drawing on the results of the North American Proficiency Testing program. Pearson’s correlation coefficients exceeding 0.60 indicated strong relationships between sensor variables and field-measured soil properties, topographic parameters and shallow ECa sensor variables, allowing effective predictions of several soil chemical properties (i.e., SOM, P, and CEC).Lastly, supervised machine learning models drawing on high-density information from multiple sensors (PSS and RS) operating at different geospatial scales, were used to generate thematic soil maps for an agricultural field in Ontario, Canada. A random forest (RF) regression model delineated the complex hierarchical relationships existing among the sensor variables and evaluated prediction efficiencies for multiple soil nutrients. The reduction of variables based on their relative importance and parameter optimization (i.e., by defining the number of trees) of the regression forest improved the predictive accuracy for nine soil properties at the cross-validation stage. The best prediction capacity has been achieved for soil pH, K, and Zn (R2 ≥ 0.80)

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.