Chapter 7: General Discussion
7.2 Satellite remote sensing application potential
In Chapter 5, 9 of the 10 VIs calculated based on the simulated wavebands of FORMOSAT-2 satellite were the same as those derived from actual satellite remote sensing images included in Chapter 4. For both cases, the CI, RVI (RVI3 in Chapter 4, RVI in Chapter 5), MCARI, TCARI, TVI, GNDVI (NDVI2 in Chapter 4, GNDVI in Chapter 5), OSAVI, and NDVI (NDVI3 in Chapter 4, NDVI in Chapter 5) were very stable in estimating the
AGB (R2 = 0.87-0.90 for actual FORMOSAT-2 data, R2 = 0.60-0.82 for simulated
wavebands) (Tables 4-5, 5-5, 5-6). Among them, the CI and RVI (RVI3 in Chapter 4) were the VIs that explained the most variation in both cases. And they were also the best and most stable VIs for estimating PNU (Tables 4-5, 5-5, 5-6). The performance of MCARI and TCARI closely followed the CI and RVI. The MCARI and TCARI are highly related to leaf chlorophyll variation and LAI (Daughtry et al., 2000; Haboudane et al., 2004). The OSAVI and RVI can weaken the effects of background reflectance, while the MCARI and CI can enhance the response to chlorophyll concentrations in addition to weakening the background noise and avoiding the saturation effect because of their linearity with chlorophyll content
(Daughtry et al., 2000; Gitelson et al., 2003, 2005). Previous studies have also observed high
linear correlations between the MCARI and N indicators such as AGB (R2 = 0.68-0.79) and
PNU (R2 = 0.83) (Cao et al., 2013; Gnyp et al., 2014). The wavebands of FORMOSAT-2
include the traditional blue (450-520 nm), green (520-600 nm), red (630-690 nm), and near- infrared (760-900 nm), which are the commonly used wavebands in many operational satellites (Table 2-2). Therefore, VIs based on the traditional bands have high potential in applications over large-area.
Crop growth stages not only affect the relationships between VIs and N nutrition indicators but also affect the selection of sensitive bands (Li et al., 2010). For the estimation
of AGB in Chapter 5, the R2 of the estimation models across the growth stages were very
high (> 0.80). However, the root mean square error (RMSE) and the relative error (REr)
obtained in the validation data using the models of across stages were greater than those based on each growth stage (PI and SE). Similar results were observed for the PNU estimation. This is because the clusters of different growth stages in the scatter plot were not evenly distributed on both sides of the regression curve, which led to a high coefficient of
determination of the fitted regression model, but higher RMSE and REr were obtained for
validation (Gnyp, 2014). Considering precision N management is often only for critical key growth stages, estimating agronomic parameters at different growth stages are important.
However, at the early crop growth stage, the performance of VIs based on the visible wavebands was poorer than that based on the red edge waveband for the estimation of N nutrition indicators (Tables 5-5, 5-6). The small vegetation coverage at this growth stage is the dominant factor. The red edge waveband has been shown to be insensitive to background effects (Zarco-Tejada et al. 2004). The results of Chapter 5 show that the VIs based on the red edge waveband can increase the coefficient of determination by more than 20% when estimating AGB and PNU at the early stages (Tables 5-5, 5-6). Previous studies showed that red edge reflectance was highly correlated with chlorophyll content (Cho & Skidmore, 2006; Clevers et al., 2002). This is because red edge position changes with the chlorophyll content (Buschmann & Nagel, 1993; Dawson & Curran, 1998) and the spectral characteristic of red edge band mainly account for N and chlorophyll content, while the visible reflectance is affected by the spectral features of multiple pigments (Haboudane et al., 2004; Hansen & Schjoerring, 2003). Chlorophyll absorbs more in the red band than in the red edge band, which is an important reason for the red edge band-VIs to be unsaturated under high chlorophyll concentrations (Gitelson & Merzlyak, 1996). It has been reported that the red edge-VIs improved the estimations of N nutrition indicators in many studies (Wu et al., 2008; Li et al., 2014a; Cao et al., 2013; Dong et al., 2015). In this study, the estimation of NNI was also improved by introducing the red edge-VIs, and the SMLR analysis proved that the red edge band was the most important one for NNI estimation except for the near-infrared band (760-900 nm) (Table 5-7). However, it was difficult to estimate PNC in the early stage of
crop growth based on satellite remote sensing according to our results. Eitel et al. (2007) also showed the difficulty of estimating leaf N concentration by using simulated RapiedEye bands.
Numerous studies have shown that high spatial and temporal satellite and aerial imagery data are useful for precision N management (Bausch & Khosla, 2010; Nigon et al., 2014; Magney et al., 2017). The advantage of satellite imagery to quickly acquire large-area data is still unmatched by drone platforms or ground-based systems. The low-cost and maintenance-free means based on satellite and aerial imagery are very helpful for accurate agricultural decision-making. Providing spatial distribution maps of in-season N uptake can help develop precision fertilization and maximize the efficiency of N use (Diacono et al. 2013; Long et al. 2015). The N recommendation approach developed in this thesis requires the satellite imagery to be collected in a narrow time window, preferably one week before topdressing N application at the stem elongation stage for rice in the study region. The satellites evaluated in this thesis all have short revisit capability and high spatial resolutions. The 8-m spatial resolution of FORMOSAT-2, 5-m for RapidEye, and 2-m for WorldView-2 can meet the requirements of agricultural production models at different scales, for instance, the large-scale farming in the Sanjiang Plain of Northeast China, and the small-scale farming in other parts of China (Shen et al., 2013). The reason for using the satellites with high temporal resolution is to minimize the impact of cloudy and rainy weather. Alternatively, Radar images and unmanned aerial vehicles (UAVs) may also be used.