Effect of Spatial Filtering on Characterizing Soil Properties From Imaging Spectrometer Data

Airborne imaging spectroscopy covering wavelength range of 0.35-2.5 mu m can be used to quantify soil textural properties and chemical constituents. In this paper, we evaluate the effects of spatial resolution on the quantification of soil constituents using a lasso algorithm-based ensemble bootstrapping framework. Airborne visible infrared imaging spectrometer data collected at 7.6m resolution over Bird's Point New Madrid (BPNM) floodway in Missouri, USA, is upscaled using a spatial filter to simulate a satellite-based sensor and generate multiple coarser resolution datasets, including the originally proposed 60.8 m hyperspectral infrared imager like data. The simulated data at multiple spatial resolutions are used in an ensemble lasso algorithm-based modeling framework for developing quantitative prediction models and spatial mapping of the soil constituents. We outline an evaluation framework with a set of metrics that considers the point-scale model performance as well as the consistency of cross-scale spatial predictions. The model results demonstrate that the ensemble quantification method is scalable, and further the model structure indicates the persistence of important spectral features across spatial resolutions. The probability density functions of the constituents over the BPNM landscape show that it is similar for multiple spatial resolutions. Finally, a comparison of the model predictions with statistical central values together with the within pixel variance across fine to coarse resolutions indicate that the model accurately captures the median values of the fine subgrid that the coarse-resolution data is composed of. This study establishes the feasibility for quantifying soil constituents from space-borne hyperspectral sensors.