Mapping soil variability with hyperspectral image data

Abstract

The spatial nature of remote sensing data presents an opportunity to characterise soils in the landscape. The characteristic spectral response of soils can also provide a powerful diagnostic tool for interpreting soil variability and predicting some soil properties. This research examined the use of hyperspectral image data to understand soil variability in a natural environment without assuming a priori or expert knowledge. The study aimed to explain the surface soil complexity within a hyperspectral image scene through the isolation of soil endmembers, and to use those endmembers to map soil variability and estimate soil exposure. Spectrally distinct soil endmembers were isolated from the image with a pixel purity routine and then used in a partial unmixing process. For the validation of results, laboratory analyses were conducted on randomly collected soil samples and field estimates of soil cover were collected. The outputs from each of the endmembers show anomalous spatial distributions, indicating that different areas are being identified and mapped. However, with the laboratory data available we were unable to demonstrate that different soils had been mapped through the unmixing process, and estimates of soil abundance were poorly correlated with field estimates. We suggest that soil properties not measured in the laboratory analysis have been captured by the unmixing process. Likely properties include variations in surface condition, such as surface crusting which is influenced by management practices; variations in soil structure and moisture, due to microtopography and management; and variations in soil colour. While the results here are limited, this methodology clearly has the potential to provide valuable information about surface soil variability relevant to agronomic and environmental land and water management issues.