An assessment of background removal approaches for improved estimation of rice leaf nitrogen concentration with unmanned aerial vehicle multispectral imagery at various observation times

Highlights

• Background effect impacted leaf N concentration (LNC) estimation with UAV imagery.

• Background removal weaked sensitivity to observation time in rice LNC estimation.

• AACI_re from sunlit pixels (AACI_re-sunlit) outperformed AACI_re from all pixels.

• AACI_re-sunlit yielded higher accuracies than SAVI and the CI_re from green pixels.

Abstract

Background effect is a crucial limitation for the monitoring of leaf nitrogen concentration (LNC) in crops with unmanned aerial vehicle (UAV) multispectral imagery. Some background removal approaches have been developed for improve the estimation of LNC, but their performances are not compared in one study and it is unclear whether they are sensitive to the observation time of UAV imagery. This study evaluated three background removal approaches, i.e., the soil-adjusted vegetation index (SAVI) approach, the green pixel vegetation index approach (GPVI) and abundance adjusted vegetation index (AAVI), for estimating rice LNC from UAV-based multispectral imagery at individual and across growth stages as well as different observation times of the day. The red edge chlorophyll index (CI_re) was chosen as the common basis for the last two approaches. In particular, the AAVI approach was refined with a higher number of endmembers and automated endmember extraction, and further evaluated for assessing the effect of separating sunlit components from shaded components of the canopy.

Our results demonstrated that the vegetation indices (VIs) for off-noon observation times showed better relationships with LNC than those for noon at individual and across growth stages. Compared to both SAVI and CI_re-green, the AACI_re for all pixels (AACI_re-all) exhibited the weakest sensitivity to observation time and yielded the best relationships for single-stage (jointing: r²=0.70, booting: r²=0.76, heading: r²=0.70) and across-stage (r²=0.66) models. Among the AAVIs derived from three categories of pixels, the AACI_re-sunlit (R² =0.90, RMSE=0.17%, Bias=0.03%) outperformed AACI_re-all (R² =0.85, RMSE=0.23%, Bias=0.08%) and then AACI_re-shaded (R² =0.38, RMSE=0.49%, Bias=0.40%) remarkably for the estimation accuracy of LNC. This study suggests that the refined AAVI approach has great value in reducing the background effect for more accurate monitoring of growth parameters and could be extended to other crops and regions for improved precision crop management and field-based high-throughput phenotyping.

Introduction

Rice (Oryza sativa L.) is one of the major cereal crops in the world, accounting for 29% of the cultivated area in China (Xiao, 2003). Its yield is not only closely related to international grain trade but also fundamental to global food security, especially in developing countries (Normile, 2008, Zhao et al., 2013, Zheng et al., 2019). Nitrogen (N) fertilizer, one of the most direct N nutrient sources in crop photosynthetic activity and productivity, plays a crucial role in improving rice yield and grain quality (Ju et al., 2009, Vitousek et al., 2009). However, excessive N fertilizer application will break the balance between supply and demand (Miao et al., 2011) and lead to a series of environmental problems such as water pollution (Chen et al., 2005) and the greenhouse effect (Sehy et al., 2003). Furthermore, the N status of rice is a crucial indicator to evaluate whether the N fertilizer application is reasonable (Miao et al., 2011). As a consequence, it is significant to monitor rice N status in order to achieve high yields and improve the N use efficiency (Shibayama and Akiyama, 1991, Zhang et al., 2011).

At present, leaf nitrogen concentration (LNC) has been commonly used as a crucial indicator for monitoring crop N status (Zhu et al., 2007). Furthermore, LNC estimation for critical growth stages is not only an important evidence for agronomists to make fertilization recommendations, but also an essential indicator to predict grain protein content (Wang et al., 2004). Compared to traditional measuring methods in the lab, remote sensing has been widely used to monitor crop LNC due to its simplicity and efficiency (Li et al., 2018, Zhou et al., 2018). It is a common practice to estimate LNC with the chlorophyll sensitive vegetation indices comprising near-infrared (NIR) and red edge bands based on the close linkage between LNC and leaf chlorophyll content (LCC) (Cheng et al., 2018, Zhou et al., 2018, Lu et al., 2019). Various remote sensing platforms have been used to monitor crop LNC, such as near-surface hyperspectral imaging systems (Tian et al., 2014, Zhou et al., 2018), unmanned aerial vehicles (UAV) (Lu et al., 2019, Wang et al., 2021), and satellites (Mutanga et al., 2015, Omer et al., 2017). Nevertheless, most of them pay more attention to the use of near-surface hyperspectral images and satellite images given the high spectral resolution or the wide field of view (Tian et al., 2014, Omer et al., 2017). Only a few studies have demonstrated that UAV platforms can be used to monitor LNC because of their low cost and ease of operation (Lu et al., 2019, Osco et al., 2020, Wang et al., 2021). However, these studies did not consistently consider the background effect (Table 1), which represents the impact of soil and water in the field on the canopy reflectance spectra because of the spectral mixing of plants and background materials in the field of view (Yao et al., 2014, Zhou et al., 2017, Zhou et al., 2018). Therefore, further research is needed to determine the necessity of background removal when estimating the LNC from UAV imagery.

The plant signal could be weakened because of the mixture of background signals in canopy reflectance, which often lead to worsened estimation of crop biochemical parameters. In this regard, some approaches were proposed to eliminate the background effect for improved monitoring of crop growth status (Gilabert et al., 2002, Jay et al., 2017a, Wang et al., 2021), such as the use of soil-adjusted vegetation indices (Gilabert et al., 2002) and extraction of green pixels for refining vegetation indices (Zhou et al., 2017, Zhou et al., 2018). Recently, (Wang et al., 2021) proposed a new approach to reducing the background effect by integrating a traditional VI and vegetation abundance through the use of linear spectral mixture analysis (LSMA). Since these approaches were used separately in different studies, it is unclear which one is the most effective and efficient to eliminate the background effect caused by soil and water, especially for the complex paddy fields. Therefore, it is beneficial to assess these approaches in reducing the background effect and determine the most useful approach for the estimation of rice LNC.

In addition, the studies on background removal paid more attention to the early growth stages (e.g., tillering and jointing) than the later growth stages (e.g., booting and heading) (Yao et al., 2014, Prudnikova et al., 2019, Wang et al., 2021). When the canopies become denser at later stages (e.g., booting and heading), the canopy reflectance is also influenced by the shadows cast by the leaves or stems in the upper layers on those in the lower layers (Jay et al., 2017a, Jay et al., 2019). Many studies have suggested that the sunlit and shaded components of the canopy should be distinguished because of the negative influence of shaded-leaf components on canopy reflectance when estimating N-related leaf traits (Zarco-Tejada et al., 2004, Jay et al., 2017a, Jay et al., 2017b). In contrast, a few studies have demonstrated that it is unnecessary to distinguish the sunlit and shaded components of the canopy in the near-ground hyperspectral images with an ultra-high spatial resolution (~ 1 mm) (Zhou et al., 2018, Jiang et al., 2021). Therefore, it remains unclear whether it is necessary to distinguish sunlit and shaded components for both early and later growth stages. It is also worth investigating whether the inclusion of shaded-leaf pixels derived from coarser-resolution UAV imagery (~ 5 cm) could affect the performance of background removal methods.

As an important factor affecting the sun-target-sensor geometry, observation time is critical to the fraction of sunlit components in the canopy and the estimation of crop biochemical parameters (Li et al., 2020). However, estimation of LNC from UAV-based multispectral imagery at multiple observation times is poorly understood. This is an important consideration because UAV flights can be completed within a short time but have to be in an appropriate time window with ideal weather conditions. Although some studies indicated the acquisition of UAV images at noon (12:00 h local time) (Jay et al., 2019, Jiang et al., 2021), others used various time windows within the period of 10:00–14:00 local time due to the stronger sunspots at noon (Liu et al., 2017, Zheng et al., 2018, Li et al., 2019, Fu et al., 2020, Osco et al., 2020, Lu et al., 2021) (Table 1). As a consequence, how the noon UAV imagery compares to off-noon UAV imagery for LNC estimation remains a valuable question to be answered. A recent study by Li et al. (2020) had demonstrated that the off-noon observation (15:00 local time) could yield better performance in LCC estimation as compared to the observation of noon (12:00 local time) at the jointing stage of rice and winter wheat. However, this task focused only on the early growth stages and it is uncertain whether their conclusion still holds for the later growth stages. Therefore, it is valuable for precision agriculture to assess the sensitivity of background removal approaches to observation time in LNC estimation.

Therefore, we compared the approaches adjusted for background removal and the traditional method without adjustment in the ability of LNC estimation for UAV-based multispectral imagery for individual and across growth stages. We also assessed the sensitivity of the background removal approaches to observation time and growth stage. The specific objectives of this study were: (1) to investigate the sensitivity of LNC ~ VI relationships for three background removal approaches to observation time, (2) to determine the best background removal approach for improving the accuracy in LNC estimation, and (3) to evaluate the effect of sunlit and shaded components on the abundance adjustment approach.