Clay content mapping from airborne hyperspectral Vis-NIR data by transferring a laboratory regression model

Abstract Visible Near-Infrared (Vis-NIR, 400–2500 nm) hyperspectral imaging has proven to be a valuable tool for mapping soil properties over bare soils. To date, most predictive models presented in literature, have been built from calibration databases made up of both Vis-NIR imaging spectra (predictor variables) and soil properties (response variables). Nevertheless, the constitution of such calibration databases is costly both in terms of time and money, as this implies soil sampling over hundreds of pixels of bare soils and physico-chemical soil analysis of these samples. We propose to test the transfer of predictive models built from laboratory database, to Vis-NIR hyperspectral image. Four transfer methods were tested (model updating, Repfile, Transfer by Orthogonal Projection TOP and Piecewise Direct Standardization PDS). Their respective performances were evaluated and compared to those obtained using two other predictive models. The first predictive model was obtained by directly applying a partial least square regression (PLSR) calibrated in laboratory, onto the hyperspectral image. The second model was obtained by applying a PLSR built from Vis-NIR spectra extracted from the hyperspectral image, onto the hyperspectral image. The transferred models are based on the use of standards. Standards were selected in order to (i) take into account or (ii) ignore field soil stratification knowledge. Moreover, the impact of the number of standards used in the transferred models was studied. This approach was tested for clay content prediction over the La Peyne Valley area in Southern France (24 km2, in a Mediterranean context) covered by HyMap Vis-NIR airborne data. The results show that 1) applying PLSR model from laboratory spectra to hyperspectral image without calibration transfer, provides poor prediction performances (R2val = 0.41, RMSEP = 103 g/kg), 2) transferring this model from laboratory spectra to hyperspectral image offered good prediction performances (R2test median above 0.5 and RMSEP median below 70 g/kg), whatever the transfer method used and from only 15 used standards, and 3) taking into account field soil stratification knowledge in the standards selection improves prediction performances (R2test median above 0.6 and RMSEP median below 60 g/kg) whatever the transfer method and from only 10 used standards. Finally, transferring models from the laboratory to a hyperspectral image gave better mapping performances than those obtained using PLSR models built from airborne spectra.