Partial Least Squares Modeling Of Hyperspectral Remote Sensing For Mapping Agricultural Soil Properties

As a common boundary where atmosphere, hydrosphere, geosphere and biosphere interact, soil plays an important role in the transfer of materials and energy between these Earth's spheres and quantitative information of soil properties are often required for environmental monitoring, modeling and precision agriculture practicing. For example, as the largest carbon stock of continental biosphere, soil organic matter (SOM) is related to the size and capacity of soil microbial population and controls soil structural stability. Effective management of soil carbon (C) reservoir can help to mitigate greenhouse gases (Mondini and Sequi, 2008). Soil moisture is related to energy exchange between soil and atmosphere and favorable to N2O and CH4 production; both soil moisture and clay content are important for plant growth and soil quality. Soil nutrient conditions drive the rate of fertilizer applications that is directly related to N2O emission and soil capacity to consume atmosphere CH4. Phosphorus (P) has been identified as a pollutant carried into water bodies from agriculture land where excess fertilizers are applied. With the increasing requirement for quantitative soil information, remote sensing has been considered as a promising tool for rapidly quantifying single or multiple soil properties. Numerous studies show that different soil properties can be estimated with laboratory measured and simulated hyperspectral data (Ben-Dor and Banin, 1990; Ben-Dor and Banin, 1994; Palacious-Orueta and Ustin, 1998). The success of these laboratory spectra-based studies naturally lead to the exploration of imaging spectroscopy for characterizing soil properties on a large scale because imaging spectroscopy not only has the capability of acquiring spectral information at several tens of hundreds spectral bands as laboratory spectroscopy does but also provides a synoptic view which cannot be achieved by laboratory spectroscopy. In contrast to a large number of airborne hyperspectral sensor applications, few attempts have been made to map soil properties with satellite hyperspectral imagery. To extract the soil information using hyperspectral data with hundreds to thousands bands, a feasible statistic method need to be applied to analyze the relationship between these two parameters. Partial least squares regression (PLS) is full-spectrum methods and commonly used to quantify hyperspectral soil data. This study aims to evaluate the potential of Hyperion imagery for estimating soil moisture content (SMC), total C, total nitrogen (N), total P, SOM, and soil clay content and to compare the estimates of these soil properties with Hyperion image spectra to those resulting from laboratory measured spectra in order to address the effects of spectral and spatial resolutions on the estimation of soil properties. The effectiveness of hyperspectral reflectance data coupled with the partial least squares (PLS) regression method for mapping agricultural soil properties is examined in this study.Soil samples were collected from central Indiana, California, USA and Castilla-La Mancha (CLM), Spain. For the samples collected in central Indiana, Hyperion images covering the Cicero Creek reservoir of central Indiana were used to estimate soil properties including soil moisture, soil organic matter (SOM), total carbon (C), total phosphorus (P), total nitrogen (N) and clay content. Two scenes of the Hyperion images were acquired, calibrated and georeferenced, and image spectra for the locations of soil samples collected in field were extracted from the calibrated and corrected images. To examine the performance of partial least square (PLS) regression for estimating major soil properties mentioned above from Hyperion data, the PLS results for Hyperion image spectra were compared with laboratory measured spectra. The laboratory spectra across the 0.4-2.5nm region (2151 bands) were collected for moist field soil samples and then resampled to 165 bands matching the spectral characteristics of Hyperion data. To conduct these comparisons, several statics were used including the coefficient of determination (R2) and RPD (the ratio of standard deviation of sample chemical concentration to root mean square error). PLS was conducted in two phases: phase-1 where all samples were used in a calibration step to determine outliers and calibrated again after outlier removal, and phase-2 where the outlier removed dataset were split into two subsets for calibration and validation. Based on R2 and RPD values, the results for the phase-1 calibration indicate that PLS can estimate some of soil properties from Hyperion spectra. In phase 2, PLS can predict most of soil properties satisfactory with Hyperion reflectance spectra except for moisture, SOM, and clay content. The higher spectra resolution increased prediction accuracy for most of soil properties except for moisture and total P. Prediction accuracy dropped to unreliable using Hyperion reflectance spectra for moisture and clay content maybe due to lower signal-to-noise ratio of Hyperion spectra and the Hyperion spectral resolution of 30 m. Spectral absorbance used to build PLS model to remove the nonlinearity on soil spectra could improve prediction accuracy for soil moisture with laboratory data, SOM with hyperspectral data and total C with resampled laboratory and Hyperion data. PLS generated the poorest results for estimating clay content especially with Hyperion data. This primarily originates from the interfering effects of soil moisture on the clay absorption feature and a low signal-to-noise ratio of Hyperion data in the shortwave infrared region. For the samples collected in California and CLM, Spain, carbonate and clay concentrations were analyzed after oven dry soil samples, but laboratory spectra were measured for soil samples at sequential moisture levels. Carbonate, clay and SMC were estimated using PLS. PLS had high performance for estimating carbonate and SM, but poor estimation for clay due to band overlapping.In the future, we want to use satellite hyperspectral data with a signal-to-noise ratio equivalent to those for HyMap and AVIRIS to improve the estimation of clay content from reflectance. The existing drawbacks in empirical models limited the spatial and temporal transferability. We will collect more soil samples to acquire wider range samples to improve the feasibility of PLS models predicting soil properties. We also want to develop a non-linear PLS models and testing them with Hyperion, HyMap and AVIRIS images.