Shining Light on Digital Agriculture: Linking Soil NIR measurements, Fertility and Crop Yields
Date: August 1, 2024
Term: 4 years, ending August 2024
Status: Complete
Researcher(s): Derek Peak, U of S
SaskCanola Investment: $103,535
Total Project Cost: $384,433
Funding Partners: ADF, Sask Wheat
Grower Benefits
Fourier Transform Near Infrared (FT-NIR) is a technology that is used to analyze the composition of substances without damaging them. The Fourier Transform refers to a mathematical process that helps convert the raw data into a form that’s easy to interpret.
Fourier Transform Near Infrared (FT-NIR) spectroscopic tools have increased in performance and decreased in price to be feasible for soil fertility measurements.
Results from FT-NIR were suitable for general soil health parameters such as soil organic carbon (SOC), total nitrogen (N), and moisture. Performance for plant available nutrients remains poor.
Models for SK soil properties worked best when using continuous wavelet transformation for data processing machine learning approaches trained with at least 4-500 samples.
Project Summary
Current commercial soil testing relies upon physically removing soil from fields and shipping it to centralized labs for analysis through wet chemistry. The soil is shipped, dried, sifted and then analyzed through chemical treatment. This process is repeated for every sample shipped. Fields are sampled repeatedly – every few years at a minimum, and every field is sampled in multiple locations. This process takes significant time and money as shipping soil is not efficient.
Soil scientists and agronomists have long sought direct measurement of soil fertility in the field, and many trials have been conducted on using near infrared (NIR) spectroscopy as a method to measure a wide range of soil chemical properties in the laboratory and in the field.
Spectral-based NIR sensing systems have the potential to reduce per-sample analysis costs by more than 90%, while producing near-instantaneous results in the field. Coupling such an increase in the efficiency of soil sampling could lead to much higher spatial resolution and would supply vital missing information for digital soil mapping, variable rate fertilizer application, and agriculture in general. However, many challenges have stymied efforts in successfully replacing lab-based soil analysis with on-site NIR. Recent improvements and miniaturization of interferometers have addressed most hardware issues such as calibration and repeatability of measurements, and software advances in data management/visualization, machine learning, and statistical approaches have largely solved challenges with using NIR as an analytical tool.
A commercial soil testing laboratory in western Canada is already collecting FT-NIR spectra of all dried soil samples and using internal calibrations to estimate soil nutrient levels using spectra. However, there are some remaining research challenges that remain to scale out on-site NIR to widespread commercial agriculture. One must either be able (a) to correct field spectra for things that can affect NIR signal (moisture, soil contact, texture, temperature) to use calibrations developed on dried laboratory soils or else (b) develop calibrations in the field directly. This project focuses on these challenges.
There were several sub-projects conducted in the four years of this study. Researchers wanted to develop a SK-based NIR model and compare SOC lab analysis approaches for NIR model development. They also looked at the effect of sample preparation on NIR model performance. In the final project, researchers explored optimizing soil organic carbon modelling accuracy by using spectral spiking with local soil spectral library data.
The largest research accomplishment and output in this study was the development of a usable model to predict soil properties from NIR spectra. Researchers collected agricultural soil samples from six different sites that spanned most of the soil zones of the province, and analyzed them for plant available nutrients, pH, SOC, total C, and total N as well as collecting NIR spectra. Initial observations suggested that the only soil properties that could be reliably predicted were SOC and total N, so future analysis focused upon these properties. Researchers also learned from model development that calcareous soils in the area near Rosetown were problematic to model, and so they investigated how to improve predictions in those samples with different laboratory approaches. New protocols were developed to add new sites to the existing model, and they developed a framework for how to extend the laboratory-based model to field-measured spectra.