Prediction of Topsoil Organic Carbon Using Airborne and Satellite Hyperspectral Imagery

Juanjo Peón, Carmen Recondo, Susana Fernández, Javier F. Calleja, Eduardo De Miguel, Laura Carretero
2017 Remote Sensing  
The Airborne Hyperspectral Scanner (AHS) and the Hyperion satellite hyperspectral sensors were evaluated for their ability to predict topsoil organic carbon (C) in burned mountain areas of northwestern Spain slightly covered by heather vegetation. Predictive models that estimated total organic C (TOC) and oxidizable organic C (OC) content were calibrated using two datasets: a ground observation dataset with 39 topsoil samples collected in the field (for models built using AHS data), and a
more » ... t with 200 TOC/OC observations predicted by AHS (for models built using Hyperion data). For both datasets, the prediction was performed by stepwise multiple linear regression (SMLR) using reflectances and spectral indices (SI) obtained from the images, and by the widely-used partial least squares regression (PLSR) method. SMLR provided a performance comparable to or even better than PLSR, while using a lower number of channels. SMLR models for the AHS were based on a maximum of eight indices, and showed a coefficient of determination in the leave-one-out cross-validation R 2 = 0.60-0.62, while models for the Hyperion sensor showed R 2 = 0.49-0.61, using a maximum of 20 indices. Although slightly worse models were obtained for the Hyperion sensor, which was attributed to its lower signal-to-noise ratio (SNR), the prediction of TOC/OC was consistent across both sensors. The relevant wavelengths for TOC/OC predictions were the red region of the spectrum (600-700 nm), and the short wave infrared region betweeñ 2000-2250 nm. The use of SMLR and spectral indices based on reference channels at~1000 nm was suitable to quantify topsoil C, and provided an alternative to the more complex PLSR method. significant impact on the atmospheric CO 2 concentration and affect the climate [3] . The measurement of these pools at the range of spatial and temporal scales still remains a challenge [4] . Rapid and accurate methods are needed for improving the inventory, spatial distribution, and temporal dynamics of SOC [5, 6] . Conventional methods for mapping SOC stocks involve the collection and analysis of point soil samples, calibration of a spatial prediction function, and interpolation of the function over the whole study area [7] . These methods are expensive and time-consuming because of the large number of samples required to capture the high spatial variability of SOC [8] . Laboratory-based measurements of soil C using standard methods, such as dry combustion [9] and the Walkley-Black method [10], also involves intensive and costly procedures. Soil diffuse reflectance spectroscopy in the visible (VIS, 400-700 nm), near infrared (NIR, 700-1300 nm), and short-wave infrared (SWIR, 1300-2500 nm) has been demonstrated to be an alternative to traditional methods for SOC prediction. Currently, lab spectroscopy is well established as an accurate, rapid, and non-destructive technique to predict a wide range of soil properties [11] , including organic C [12] [13] [14] [15] and total C [16] [17] [18] . Lab spectroscopy can be extended to a regional scale using airborne and satellite hyperspectral sensors. However, the use of hyperspectral sensors at a remote sensing scale has several limitations, such as atmospheric absorptions, illumination variations, the low signal-to-noise ratios of the sensors, and spectral mixing [19, 20] . Due to these limitations, relatively few studies have used data from airborne and satellite hyperspectral sensors to estimate SOC and organic matter, either at the within-field scale or at the regional scale. Several hyperspectral airborne sensors were tested to predict SOC and organic matter over agricultural fields or semi-arid areas, including HyMap [21-26], Airborne Hyperspectral Scanner (AHS) [8,27-30], Compact Airborne Spectrographic Imager (CASI) [31,32], Airborne Visible and Near-Infrared (AVNIR) [33], DAIS-7915 [34], and HyperSpecTIR [35]. However, very few studies used airborne hyperspectral data to estimate SOC and organic matter over areas partially covered by vegetation, in either agricultural areas [36, 37] or burned areas with a partial vegetation cover resulting from post-fire regeneration [38] . Satellite hyperspectral data was seldom used to estimate SOC, and only one satellite sensor, Hyperion, was used to predict SOC and organic matter over agricultural fields and bare soils [39] [40] [41] . Hyperion data was also used to estimate SOC and organic matter over maize crops [42] , and over forests, pastures, and agricultural fields with inaccurate results [43] . The prediction of SOC using spectroscopy is usually based on several multivariate statistical techniques or data mining algorithms, such as multiple linear regression, stepwise multiple linear regression [44, 45] , principal component regression [46, 47] , regression trees [48], support vector machine regression [18] , and artificial neural networks [49] . Partial least squares regression (PLSR) is one the most widely-used techniques, mainly because it handles multicollinearity in the reflectance spectra, and it is robust in terms of data noise [11] . However, an important drawback is that PLSR does not provide a quantitative explanation for the relationship between predictor variables and response variables [50] , and it is complex to transfer models from one sensor to another [51] . Linear regression methods using spectral indices are less sophisticated than PLSR, easier to transfer among sensors, and based on the physical analysis of spectral reflectance, so they are used as an alternative modeling method to PLSR [50, 51] . So far, the estimation of SOC using airborne and satellite hyperspectral sensors was mainly restricted to small agricultural or bare soil areas, and it still remains in the testing phases. Furthermore, the calibration technique used for SOC prediction in most of the previous studies was the PLSR method, which has several drawbacks related to the physical interpretation of the results and the complexity of transferring the models from one sensor to another. The aim of this work was to assess the capability of two hyperspectral sensors, the Airborne Hyperspectral Scanner (AHS) and the Hyperion satellite hyperspectral sensor, to estimate SOC in burned mountain areas that are slightly covered by vegetation as a result of post-fire regeneration. Remote Sens. 2017, 9, 1211 The SOC estimation was carried out using simple and multiple linear regression techniques based on image reflectance values and spectral indices (SI), which are less sophisticated than the widely-used PLSR method, and have greater potential to be transferred among sensors. Comparison of prediction performance with PLSR results allowed us to evaluate whether the proposed regression models provide an acceptable accuracy. Materials and Methods Study Area The study area is located on the western side of the Cantabrian Range (NW of Spain), and consists of a NW-SE rectangle of approximately 60 km 2 , which corresponds to the AHS image extent (Figure 1 ). The area corresponds to a mountain region with an altitude range from 400 to 1700 m above sea level, and an average slope of~25 • . The climate is Atlantic, with an average annual air temperature of 8 • C and an average annual precipitation of 1500 mm. The bedrock is very homogeneous, and is mainly composed of quartzite, sandstone, and slate. The soils are classified as Lithosols, Histosols, and Regosols, according to the World Reference Base [52] , and are sandy, shallow, and stony. We chose this region because it has been affected by frequent wildfires [53] , and soil in this area contains high amounts of organic matter [54] , which makes it well-suited for the study. As a result of the degradation caused by the repeated burning of this region, the landscape is dominated by extensive areas of heath formations, and deciduous forests, to a lesser extent [55] . The abundance of heaths within the burned slopes located in the study area varies along a N-S gradient, with higher abundance in the northern areas. The wildfire incidence map obtained with Landsat [56] data showed regions of the study area affected by up to four wildfires from 1984 to 2011. The last wildfires in the study area were recorded in 2005 (center of the study area), 2006 (south), and 2011 (north and center-south).
doi:10.3390/rs9121211 fatcat:4geyf7yqongmxe2p3zgz63zmh4