Monitoring Nitrogen Nutrition Parameters With Hyperspectral Remote Sensing In Rice

The spectral parameters and monitoring models in remote sensing studies are of key importance for information acquisition and diagnosis on crop growth status. At present, development of digital agriculture presents an urgent need for low-cost, reliable, consistent and precise techniques for monitoring nitrogen status in crop plants. Recent success of remote sensing in agricultural application has made it possible to rapidly monitor growth status and biochemical components in large-area field crops. Implication of various ground-based and space-borne remote sensed information and instruments is receiving more and more attention in existing research. This study was undertaken to make a systematic analysis on the characteristics of the reflectance spectra, compare prediction ability of nitrogen status with vegetation indices based on narrow and wide bands, and develop key spectral indices and quantitative models for nitrogen parameter estimation at ground and space levels, based on canopy hyper-spectral (multi-spectral) and image spectral reflectance of field-grown rice with varied nitrogen levels and rice varieties in different years. This would help to establish the key technology for real-time monitoring of plant nitrogen status with ground and satellite hyper-spectral sensors in rice production.Firstly, a systematic analysis was made on the characteristics of the first-derivative reflectance spectra in red edge area, and the quantitative relationship of red edge position (REP) and red edge area shape parameters to canopy leaf nitrogen concentrations under varied nitrogen rates and rice varieties in different field experiments. The results showed that spectrum in red edge area was significantly affected by different nitrogen levels and different rice varieties, and "three-peak" feature could be observed with the first derivative spectrum at about700nm,720nm and730nm bands, respectively. Traditional REP was not sensitive to canopy leaf nitrogen concentration because of the three-peak feature, but REPs based on inverted Gaussian fitting technique, linear four-point interpolation technique, linear extrapolation method and adjusted linear extrapolation method generated continuous REP data, and could be used to estimate canopy leaf nitrogen concentration. Besides, REP from a three-point Lagrangian interpolation with three first-derivatives bands (695nm, 700nm and705nm) also had a good relationship with canopy leaf nitrogen concentration. Comparison of these REP revealed that the adjusted linear extrapolation method (755FD73o+675FD700)/(FD730+FD700) had the best prediction performance on canopy leaf nitrogen concentration, with a relative simple algorithm, so it is a proper REP parameter for estimating canopy leaf nitrogen concentration in rice.In addition, the peak heights of the3peak bands changed alternatively with different nitrogen levels, so child areas and shapes surrounded by the first derivative spectra curve and x coordinate changed accordingly. Double peak symmetry (DPS) based on the ratio of2different red edge child areas divided by "peak band line", and normalized double peak symmetry (NDPS) with normalization of the2different red edge child areas were significantly related to canopy leaf nitrogen concentrations in rice. Results of model calibration and validation indicated that DPS(A675-700,A675-755) and NDPS(A675-700, A675-755), ratio and normalized difference of area in675-700nm to675-755nm red edge region, respectively, performed best in estimating canopy leaf nitrogen concentration, so these two spectral indices were good red edge area shape parameters for monitoring canopy leaf nitrogen concentration in rice. Furthermore, the spectral parameter DPS(A72o-730A700-720) was found to be a good indicator for leaf nitrogen accumulation in rice.Then, the sensitivity of reflectance spectra to canopy leaf nitrogen status was examined, and the quantitative relationships of different hyper-spectral vegetation indices to canopy leaf nitrogen concentrations were evaluated. The results showed that the reflectance of red region (665-675nm), blue region (490-500nm) and red edge region (680-760nm) was highly sensitive to canopy leaf nitrogen status. Two bands-based vegetation indices combined with550-600nm and500-550nm in green region had good relationships with canopy leaf nitrogen concentrations, and ratio index R(533,565) exhibited the best performance in all two bands vegetation indices. Yet prediction ability of nitrogen concentration was significantly improved using three bands vegetation indices. Novel three bands indices, blue nitrogen indices R434/(R496+R401) and R705/(R717+R491) were developed for monitoring canopy leaf nitrogen concentration, and the test results indicated that R434/(R496+R401) and R705/(R717+R491) had better prediction precision and suitable application than R(533,565). In addition, some reported vegetation indices also had good relationships to canopy leaf nitrogen concentrations, such as two bands indices ZM, GM-2, RI-1dB, RI-2dB, NDRE and three bands indices mND705, PRIc, with three band indices better than two band indices, although less satisfactory prediction than blue nitrogen indices and R.705/(R705+R491). Comparison of all previous indices and present indices indicated that novel blue nitrogen indices R434/(R496+R401) and R705/(R717+R491) had the best prediction capability for estimating canopy leaf nitrogen concentration in rice. Besides, RVI(827,742) was identified as a good indicator for leaf nitrogen accumulation in rice.A knowledge on the responses of reflectance spectra and vegetation indices to bandwidth can contribute to proper selection of sensitive bands and development of spectral indices for nitrogen status monitoring. Effects of different bandwidths on canopy reflectance spectra in the range of350to1000nm bands and sensitive vegetation indices were studied with simulated spectra based on ASD data, and then prediction powers of canopy nitrogen concentrations based on two different remote sensors, ASD and Cropscan were compared. The results showed that there was less effect on reflectance spectra if bandwidth was within20nm, on the contrary, when it was larger than50nm, several phenomena happened that the red vale turned shallow, green peak got lower and slope of red edge reached flat. Bandwidth had smaller impact on visible light and larger effect on near infrared light under different nitrogen levels, discrimination power of nitrogen levels based on red light decreased with reduced spectral resolution, but green light could well discriminate nitrogen levels within200nm spectral resolution. Ratio indices, normalized indices and differential indices developed from reflectance of NIR and red edge, red and green bands also decreased with reduced spectral resolution, especially when bandwidth was broader than20nm. Yet RI(NIR,Green) and NI(NIR,Green) just had similar values when bandwidth was narrower than200nm. Several sensitive vegetation indices such as R434/(R496+R401), R705/(R717+R491) and R533/R565were found to be very stable within300nm bandwidth, thus indicating a wide applicability with different remote sensors.Relationships between canopy leaf nitrogen concentration and spectral indices composed of arbitrary two bands with different spectral resolution in the region of350to1000nm were also explored. The results indicated that the band combinations with good correlation to nitrogen status were constant within10nm bandwidth, and then reduced in number. For the purpose of reliable estimation of nitrogen status, the bandwidths of vegetation index such as ND (760,710) maybe diverse. The200nm bandwidth for760nm contributed identically to nitrogen estimation, but20nm bandwidth was requested for710nm, and different vegetation indices exhibited different powers. Comparison of Cropscan and ASD sensors indicated that reflectance spectra were not only controlled by bandwidth, but also by applied nitrogen levels, cultivar types and growth stages. The ASD sensor had higher prediction power than Cropscan sensor with the same vegetation index at the growth stages with higher coverage, whereas vice versa at growth stages with lower population coverage.Further, the relationships of leaf chlorophyll status, leaf area index and leaf photosynthesis to hyper-spectral reflectance and derivative parameters were quantified for deriving monitoring models on leaf pigment status with key hyper-spectral indices in rice. The results indicated that ratio index RI(714,760), normalized index ND(543,565) and difference index DI(562,543) had good relationships with chlorophyll a or chlorophyll a+b concentrations. The first derivative spectral ratio and normalized indices R(D744,D761) and ND(D748,D761) had better relationships with chlorophyll a or chlorophyll a+b concentrations than RI(714,760), ND(543,565) and DI(562,543), and the similar results were obtained with three bands index gmND705. The correlation results showed that vegetation indices composed of spectral reflectance of bottom band in red edge area743nm and NIR band822nm were significantly related to chlorophyll density, and both RI(743,822) and ND(743,822) had negative linear relationships with canopy leaf chlorophyll density. The first derivative spectral normalized and difference indices ND(D511,D771) and DI(D549,D779) also had better relationships with chlorophyll density, although not much improved over the RI(743,822) and ND(743,822). Green modified normalized index gmND705was good indicator for canopy leaf chlorophyll concentration, and normalized index ND(743,822) for chlorophyll density estimation from comparison of determination coefficients, RMSEs and REs of monitoring models with different vegetation indices. In addition, good relationships existed between LAI and vegetation indices, the correlation sequence of LAI to different index types was DVI>RVI>NDVI which consisted of spectral reflectance or the first derivative spectra. The best vegetation index for LAI monitoring were found to be the difference index of850nm to760nm DVI(854,760) and the first derivative difference index of676nm to778nm DVI(D676,D778), respectively. The vegetation index composed of spectral reflectance was better than that of the first derivative spectra for LAI monitoring in rice. Testing of the monitoring models with independent dataset also proved that the spectral index of DVI(854,760) gave accurate LAI estimation, so can be considered as a good indicator for LAI monitoring in rice. In addition, ratio index R(810,680) can be used to monitor leaf photosynthetic characteristics at different growth stages of rice.Finally, an integrative analysis was made on the characteristics of reflectance spectra of ASD, Hyperion, TM and ALOS, and on the quantitative relationships of spectral parameters to canopy leaf nitrogen concentrations under different nitrogen levels. The results showed that real spectral reflectance in rice was significantly influenced by atmosphere, which caused the reflectance of VIS to increase and NIR to decrease. Atmospheric correction model FLAASH and empirical line calibration were good methods for Hyperion, TM and ALOS, respectively. The hyper-spectral parameters for canopy leaf nitrogen estimation were essentially consistent at ground and space levels, and green modified normalized difference index gmND(760,710), adjusted linear extrapolation red edge position, RVI(884,690) and NDVI(884,690) were good predictors for canopy leaf nitrogen concentrations in rice. As compared to Hyperion data, hyper-spectral parameters such as REP could not be extracted from TM and ALOS data, but blue modified normalized vegetation index (RNIR-RR)/(RNIR+RR-2×RB) composed of reflectance of blue, red and NIR was better than NDVI(NIR,R) for canopy leaf nitrogen concentration evaluation with the same prediction precision. In addition, RI(830,742) and ND(NIR,Green) derived from Hyperion and TM or ALOS respectively were two good parameters for estimation of canopy nitrogen accumulation at ground or space level. Since ALOS had the similar ability of estimating canopy leaf nitrogen concentration as TM with higher spatial resolution, it could be quite promising for plant growth monitoring in the future.