A 2-year field experiment was carried out from 2021 to 2022 at the Qiyuan farm site (30°100E, 119°480N), Hangzhou City, China. The experimental site has a typical subtropical monsoon climate condition and is characterized by abundant precipitation with mild seasonal temperature variation. The yearly mean annual temperature and total precipitation are 18.7°C and 1,930 mm, respectively. The predominant soil is characterized as loamy, having a pH of 7.9, total N content of 1.45 g/kg, hydrolytic N content of 199 mg/kg, and organic matter of 24.1 g/kg.

In 2021, Choy Sum was transplanted on April 15 and harvested on May 24. In 2022, Choy Sum was transplanted on April 1 and harvested on April 25. The average temperature was 22.9°C and the total precipitation was 147.0 mm in the growth season of 2021, while the average temperature was 18.1°C and the total precipitation was 121.7 mm in the growth season of 2022. Pumpkin and sweet potato were the previous crops in both seasons. Each plot had a size of 5 m × 5 m. The plant density was 0.15 m × 0.15 m and 0.3 m × 0.3 m in 2021 and 2022, respectively. Six treatments were included in each experiment differentiated by varying N rates of 0 kg/ha, 25 kg/ha, 50 kg/ha, 100 kg/ha, 150 kg/ha, and 200 kg/ha (defined as N₀, N₂₅, N₅₀, N₁₀₀, N₁₅₀, and N₂₀₀, respectively). The experimental design utilized in the field was a completely randomized block design, which consisted of three replications. Urea was used to apply N fertilizer at transplantation, while 30 kg/ha of phosphorus (P) and 90 kg/ha of potassium (K) were applied simultaneously. The administration of irrigation, pest control, and disease management were conducted in accordance with regional best practices.

A total of 90 and 54 samples were collected in 2021 and 2022, respectively (Table 1). Each plot had a sub-plot of 0.5 m × 0.7 m marked for sampling, and plants within the sub-plot area were collected and stored for subsequent analysis. Digital images and spectral measurements were taken prior to destructive sampling. All samples were desiccated in the oven at 105° C for 30 minutes and then at 75° C for 48 hours to measure dry matter, which refers to the dry weight of above-ground plant material per unit area. Subsequently, plant N concentration (PNC, %) of dried samples was determined by the Kjeldahl method (KDN-B, Shanghai Xinjia Electronic, Co., Ltd., Shanghai, China).

A hand-held field spectrometer (RapidScan CS-45; Holland Scientific, Lincoln, NE, USA) was used to measure canopy spectral reflectance at 670 nm (defined as Red), 730 nm (defined as RedEdge), and 780 nm (defined as near infrared (NIR)) wavelength and provides normalized difference red edge (NDRE) and normalized difference vegetation index (NDVI) (Aranguren et al., 2019). For each plot, a representative row was selected for spectral data measurement. The instrument was maintained approximately 1 m above the crop canopy during the scanning procedure. The four most commonly used vegetation indices named NDRE, NDVI, red edge ratio vegetation index (RERVI) (Erdle et al., 2011), and ratio vegetation index (RVI) (Zhou et al., 2017) were calculated as follows.

A digital camera (D5600, Nikon, Tokyo, Japan) was used to capture crop images, which were held above the crop canopy at a distance of 1 m. Images were converted to HSI three-color space (hue, saturation, and intensity channels) using Matlab (The MathWorks, Inc., Natick, MA, USA) programming. The threshold of H values was used to generate binary pictures. Those pixels with H values between 0.2 and 0.4 were given the value 1 (green vegetation), while other pixels without vegetation were put to 0 (soil background). Canopy coverage was calculated as the proportion of the number of pixels with a value of 1 to the whole pixel number in the binary image.

Photosynthetic active radiation was measured by a light quantum sensor (model Li-190R, Li-Cor Inc., Lincoln, NE, USA), which was installed near the experimental site. The radiation sensor was connected to a data logger (model CR3000 Series), which was used to record incident photosynthetic active radiation every 10 minutes. Daily intercepted photosynthetic active par (Ipar) and accumulated Ipar of plant canopy were calculated as follows:

where I₀ is photosynthetic active radiation retrieved directly from the sensor. Fipar is the fraction of photosynthetic active radiation, which was substituted by observed canopy coverage in this study. A i p a r is accumulated intercepted photosynthetic active radiation, with t as the growth period under consideration.

The coefficient of determination (R²) was utilized to assess the general fit of the regression equation. Moreover, root mean square error (RMSE) indicates the performance of regression models. They were calculated as follows:

where N is the number of observations, P i   is the observed value, P ¯ is the mean of observed values, and O i is the predicted value.

The Tukey–Kramer method was applied to test the significant differences of PNC, DM, and RUE under different N treatments. The nominal alpha value of 0.05 was applied to determine R² and significance levels. In this study, OriginPro 2023 (OriginLab Corp., Northampton, MA, USA) was used to draw all graphs, and SPSS 14.0 software (SPSS Inc., Chicago, IL, USA) was used to perform statistical analysis.

Table 2 displays the variation in PNC at different growth stages. In 2021, the PNC ranged from 3.02% to 5.71%, while in 2022, it ranged from 3.14% to 4.99%. In both seasons, the PNC of Choy Sum declined with the progress of crop growth. In 2021, the PNC of N₀ and N₂₅ treatments was significantly lower than that of the other treatments (N₁₀₀, N₁₅₀, and N₂₀₀). The results in 2022 showed a similar pattern to 2021, although PNC of N₂₅ only showed a significantly lower value until 25 days after transplanting.

The dry matter of Choy Sum varied from 4.30 g/m² to 93.12 g/m² in 2021 and varied from 3.87 g/m² to 20.57 g/m² in 2022 (Table 3). In 2021, a notable difference was detected in the dry matter between low N treatments (N₀ and N₂₅) and high N treatments (N₅₀, N₁₀₀, N₁₅₀, and N₂₀₀) at all growth stages. In 2022, there was also an obvious difference in the amount of dry matter between high N treatments (N₁₀₀, N₁₅₀, and N₂₀₀) and low N treatments (N₀ and N₂₅) at 16, 20, and 25 days after transplanting.

Dry matter was estimated by two different models in this study, i.e., spectral vegetation index-based and Aipar-based models. Dry matter and vegetation indices were observed to be closely related. Their relationship can be best fitted by exponential models (Figure 1), while the performances of linear models were relatively worse than those of the exponential ones. The coefficient of determination of all exponential models had considerable R² of 0.74, 0.76, 0.75, and 0.76 for NDRE, NDVI, RERVI, and RVI, respectively. In contrast, the correlation between dry matter and Aipar was significantly linear (R² = 0.82).

A linear relationship between DM and Aipar was detected across different N rates in both 2021 and 2022 (Figure 2; Table 4). The linear relationship was derived from the equation of DM = RUE × Aipar, where RUE was treated as the slope of the linear function, and a comparison of RUE of different N treatments was also conducted. In both 2021 and 2022, the RUE of N₀ and N₂₅₀ was substantially lower than the RUE of other treatments. N₂₀₀ obtained the highest RUE of 0.21 in 2021, while N₁₀₀ obtained the highest RUE of 0.11 in 2022. Significant differences in RUE between N₁₀₀, N₁₅₀, and N₂₀₀ were not detected. The R² values of all regression models were higher than 0.95.

In this study, both Aipar-based and spectral reflectance-based models were adopted to estimate the dry matter yield of Choy Sum, and the performances of these two types of models were comprehensively compared. Four commonly used vegetation indices were extracted from the active field spectrometer to build estimation models of Choy Sum yield.

The relationships between vegetation indices and dry matter yield were best fitted by exponential models, which outperformed the linear ones, suggesting saturation problems occurred when dry matter approached a certain high level (Figure 1). This phenomenon agreed with the findings of other studies, where saturation often occurred for high biomass or LAI values in the late growth stage. For instance, it occurs when biomass reached 3 t/ha or NDVI approached 0.95 for rice (Gnyp et al., 2014), or NDVI reached 0.8 for wheat (Erdle et al., 2011), while it occurred when NDVI was approximately 0.8 for Choy Sum in this study. Nonetheless, dry matter can still be satisfactorily estimated by vegetation index-based models over the two seasons (Figure 1).

The saturation issue could be well addressed using the Aipar-based model, as Aipar increased proportionally to dry matter, even when the canopy approached a high level (Figure 1); this is due to the fact that the parameter of Aipar integrates not only canopy coverage but also actual sun radiation. The constraint of the study is that Fipar was not directly measured; alternatively, canopy coverage was used as a substitution for Fipar. This could lead to a deviation of the true RUE values, although Haverkort reported that canopy coverage could be used as an approximation of Fipar on potato crops (Haverkort et al., 1991; Chakwizira et al., 2015). Many studies calculated Fipar from remote sensing or leaf area index since there exist close relationships and turned out to be relatively accurate (Zhou et al., 2017; Chakwizira et al., 2018; Peng et al., 2021). Nonetheless, the Fipar estimation method in this study has rapid access at a lower cost, despite some precision sacrificed.

Furthermore, since only 2 years’ data in one site were contained in this study, and there existed many factors of environmental variation, the specific model may not be so universal over a large scale of time and space but would be a similar trend.

In both 2021 and 2022, significantly lower RUE was detected in non-nitrogen (N₀) and low-nitrogen treatments (N₂₅) than in the high N treatments of N₅₀, N₁₀₀, N₁₅₀, and N₂₀₀, among which RUE showed no difference. This indicated that the accumulation of dry matter was caused by the increase of not only Aipar but also RUE. When the N rate was above 50 kg N/ha, RUE will not be a restriction factor for DM accumulation, implying DM increase was entirely caused by Aipar increase resulting from N application. Crops have different strategies to cope with N deficiency; some crops maintain chlorophyll content to keep constant radiation use efficiency, i.e., with RUE unaffected. The research in potatoes showed that RUE remained unchanged at different N levels (Millard and Marshall, 1986; Bangemann et al., 2014), as potatoes tend to reduce the leaf area instead of photosynthesis efficiency per unit leaf area when N is insufficient (Vos and van der Putten, 1998); a similar phenomenon is also demonstrated on beet crops (Chakwizira et al., 2018). In contrast, other studies showed that RUE increased significantly with N application and leveled off when the N rate reached to medium level on corn (Muchow and Davis, 1988; Kaur et al., 2012; Kar and Kumar, 2016). In the case of Choy Sum, increased leaf N content was mainly used for thylakoid accumulation in photosynthetic cells and synthesis of carboxylases, rather than increasing leaf area to capture light (Khairun et al., 2016), corroborating well with the discovery of this research. In conclusion, N application significantly increases the RUE of Choy Sum when the N rate is below 50 kg N/ha. However, beyond this threshold, further increases in the N rate did not affect RUE.