Using satellite hyperspectral imagery to map soil organic matter, total nitrogen and total phosphorus

Abstract

Up-to-date and accurate information on soil properties is important for precision farming and environmental management. The spatial information of soil properties allows adjustments of fertilizer applications to be made based on knowledge of local field conditions, thereby maximizing agricultural productivity and minimizing the risk of environmental pollution. While conventional soil sampling procedures are labor-intensive, time-consuming and expensive, remote sensing techniques provide a rapid and efficient tool for mapping soil properties. This study aimed at examining the capacity of hyperspectral reflectance data for mapping soil organic matter (SOM), total nitrogen (N) and total phosphorus (P). Soil samples collected from Eagle Creek Watershed, Cicero Creek Watershed, and Fall Creek Watershed were analyzed for organic matter content, total N and total P; their corresponding spectral reflectance was measured in the laboratory before and after oven drying and in the field using Analytical Spectral Devices spectrometer. Hyperion images for each of the watersheds were acquired, calibrated and corrected and Hyperion image spectra for individual sampled sites were extracted. These hyperspectral reflectance data were related to SOM, total N and total P concentration through partial least squares (PLS) regressions. The samples were split into two datasets: one for calibration, and the other for validation. High PLS performance was observed during the calibration for SOM and total N regardless of the type of the reflectance spectra, and for total P with Hyperion image spectra. The validation of PLS models was carried out with each type of reflectance to assess their predictive power. For laboratory reflectance spectra, PLS models of SOM and total N resulted in higher R2 values and lower RMSEP with oven-dried than those with field-moist soils. The results demonstrate that soil moisture degrades the performance of PLS in estimating soil constituents with spectral reflectance. For in-situ field spectra, PLS estimated SOM with an R2 of 0.74, N with an R2 of 0.79, and P with an R2 of 0.60. For Hyperion image spectra, PLS predictive models yielded an R2 of 0.74 between measured and predicted SOM, an R2 of 0.72 between measured and predicted total N, and an R2 of 0.67 between measured and predicted total P. These results reveal slightly decreased model performance when shifting from laboratory-measured spectra to satellite image spectra. Regardless of the spectral data, the models for estimating SOM and total N consistently outperformed those for estimating total P. These results also indicate that PLS is an effective tool for remotely estimating SOM, total N and P in agricultural soils, but more research is needed to improve the predictive power of the model when applied to satellite hyperspectral imagery.