Peanut Leaf Wilting Estimation From RGB Color Indices and Logistic Models

Introduction

Peanut is an important oil and food crop grown on 28 million hectares (ha) worldwide (FAO STAT, 2020). In the United States, peanut provides relatively high net returns for growers; it is grown annually on approximately 619,000 ha in 11 states with an average production of 4400 kg ha^–1 (US Department of Agriculture-National Agricultural Statistics Service, 2019). Biotic and abiotic stresses are major constraints to peanut production in all peanut growing regions. For example, in a subhumid region, such as Virginia and the Carolinas, the summer (May through October) precipitation is regularly between 500 and 1000 mm. Peanut grows on sandy soils, and in southeast Virginia, where peanut is grown, available water capacity in the first 25 cm of the soil profile varies from 0.10 to 0.15 cm/cm. From May to June, precipitation is intense, and consequently, biomass grows fast. July and August are the hottest months, and peanut is at the beginning flowering stage in early July when a combination of abundant biomass, intense water evaporation from sandy soils, and absence of precipitation can induce drought and wilting in 10 days. Under these conditions, moderate drought can reduce economic return by 20% and severe drought by 60% (Balota et al., 2020). In this environment, drought usually occurs in midseason, i.e., abundant flowering, pegging, and pod growth (July and August), not at terminal growth (September and October). It also happens fast, and plants do not translocate from older to younger leaves like cereals do. This is why peanut does not senesce under drought, likely severely wilts, ceases growth, and recovers at the first rain as in Figure 1. At the same time, high relative humidity and abundant morning dew are important water sources for peanut during water stress time as suggested in other sandy ecosystems (Liu et al., 2020). Low soil moisture stress during the flowering and pegging growth stages causes severe reduction of pod yield (Stansell et al., 1976; Pahalwan and Tripathi, 1984; Naveen et al., 1992; Smartt, 1994; Rucker et al., 1995; Prasad et al., 1999; Reddy et al., 2003; Shao et al., 2008). Drought stress during the pod and seed filling growth stages results in small and immature seeds with reduced germination, vigor, embryo membrane integrity, and embryo RNA content (Reddy et al., 2003). Calcium, usually applied at the beginning of pegging in early July, is a critical nutrient for peanut seed development; it needs to be dissolved in soil solution for absorption by pods and seeds in the ground. Low soil moisture during pod development results in calcium deficiency causing undeveloped seeds (or “pops”) and embryo damage (Skelton and Shear, 1971; Wright et al., 1991). Low soil moisture during seed maturation results in a decreased oleic to linoleic fatty acid ratio, which reduces the storage shelf life and nutritional qualities of peanuts (Hashim et al., 1993). Low moisture is responsible for reduced nodulation and nitrate reductase activity affecting N fixation and uptake and biomass and yield production (Lenka and Mishra, 1973; Kulkarni et al., 1988; Devries et al., 1989). Low moisture stressed peanuts are prone to Aspergillus flavus mold contamination, known to produce the carcinogenic aflatoxin (Wilson and Stansell, 1983; Sanders et al., 1993; Luis et al., 2020). When available, supplemental irrigation can ameliorate low moisture stress, but this increases input costs. For example, contacts with growers in Virginia estimate an irrigation cost of $1.24/ha and mm water (Balota, personal communication). From planting to harvest, peanut requires weekly amounts of 25–50 mm of water (Rowland et al., 2012; Putnam et al., 2014). This suggests a weekly irrigation cost of $62 per ha. In 2020, numerous peanuts fields in southeast Virginia needed irrigation every week in July to ensure adequate water supply for the crop (Balota, personal communication).

Under drought conditions, cultivars with tolerance to low moisture stress are required. Previous studies suggest that breeding selection using physiological characteristics is a better option than being selection for yield alone (Kiniry et al., 2005; Nigam et al., 2005; Nigam and Aruna, 2007; Arunyanark et al., 2008). The use of remote sensing for phenotyping is also reported (Raju et al., 2016; Reynolds and Langridge, 2016; Sreeman et al., 2018). Studies suggest that traditional breeding methods are slow, and use of remote sensing is required for faster and more accurate phenotyping (Zaman-Allah et al., 2015; Reynolds and Langridge, 2016; Sreeman et al., 2018). Aerially derived plant wilting from low-cost cameras and RGB indices could be used to select breeding lines with drought tolerance. If directly associated with yield, aerially derived wilting could be used to trigger irrigation in peanut production, similar to how remotely sensed canopy temperature has been used to trigger irrigation in Texas (Evett et al., 1996). This ensures economical water use as suggested in similar applications in Georgia (Rowland et al., 2012).

Tolerance to low soil moisture is associated with deep rooting patterns of peanut plants. Deep roots uptake water from a lower soil horizon during drought stress (Fukai and Cooper, 1995; Henry et al., 2011; Comas et al., 2013). Accurately measuring rooting depth is difficult and labor intensive with current methods (Courtois et al., 2013; Krišāns et al., 2020). Leaf wilting, however, is easily visible and represents the phenotypic expression of plants when roots cannot replenish the water lost through transpiration. Water-deficient cells have low cellular turgor pressure, causing the leaf to lose structural integrity. This causes the leaves to either fold, roll, or droop down (Blum, 2011). These symptoms constitute visual expression of wilting and can be used as a proxy for low moisture stress in plants. Studies show that reduced leaf water potential due to low moisture stress is directly related to wilting severity (Engelbrecht et al., 2007). As illustrated in Figures 1A,B, non-wilted plants have leaves facing upward with no stems or bare ground visible compared with wilted plants with folded leaves. Because, in peanut, wilting is a clear visual symptom of low moisture stress, quantifying wilting is recommended as an important step toward development of cultivars with tolerance to low soil moisture (Luis et al., 2016). Visual rating is an important tool to quantify leaf wilting in plants (Engelbrecht et al., 2007; Hamidou et al., 2012; Luis et al., 2016; Balota and Oakes, 2017; Zhou et al., 2020). Visual rating is based on morphological changes of leaves when cells become less turgid (Figure 1). However, rating thousands of plots is time-consuming and subject to human error (Milberg et al., 2008; Borra-Serrano et al., 2018). Fast and accurate methods are required to intensify phenotyping for drought tolerance in breeding programs.

Remote sensing can be used to quantify wilting based on the changes in the leaf physiology and morphology. Red, near infrared, and infrared imagery helped with development of an abundance of vegetation indices, e.g., normalized difference vegetation index (NDVI), green based NDVI (g-NDVI), red–green ratio index (RGRI), normalized green-red difference index (NGRDI), and normalized sunlit shaded index (NSSI). They were used to estimate low soil moisture stress in several agronomic and horticultural crops, including wheat (Triticum aestivum L.) (Oakes et al., 2020), sorghum [Sorghum bicolor (L.) Moench] (Sadeghpour et al., 2017), maize (Zea mays L.) (Wenting et al., 2014), barley (Hordeum vulgare L.) (Romer et al., 2012), soybean [Glycine max (L.) Merr.] (Zhou et al., 2020), potato (Solanum tuberosum L.) (Zakaluk and Ranjan, 2008), and spinach (Spinacia oleracea L.) (Raza et al., 2014). In these studies, multispectral and hyperspectral sensing from proximal and unmanned aerial vehicle (UAV) images were used. These expensive technologies require a high level of technical knowledge (Moshou et al., 2004).

Alternative methods derived from relatively low-cost RGB digital cameras were also developed for high-throughput phenotyping (HTP) of important agronomic crops but not for peanut (Diéguez-Uribeondo et al., 2003; Reyniers et al., 2004; Graeff et al., 2006; Mirik et al., 2006; Su et al., 2020; Zhou et al., 2020). Low-cost technology coupled with open-source computer software seem to offer advantages to more complex technologies when deriving vegetation indices capable of assessing the effect of abiotic stress on crop canopy (Casadesús et al., 2007; Casadesus and Villegas, 2014). Vegetation indices can be derived from RGB color space indices, which represent international standards for color perception by the human eye and were adopted by the Commission Internationale de l’Eclairage (CIE) in 1976 (Yam and Papadakis, 2004; Trussell et al., 2005; Casadesús et al., 2007; Liu et al., 2011; Kipp et al., 2014; Zhou et al., 2015). Several studies show that RGB color space indices, such as hue, a^∗, u^∗, and other derived indices, such as green area (GA) and the normalized difference CIELab index (NDlab), outperformed spectral indices such as NGRDI, NDVI, and gNDVI, in predicting yield of wheat and maize more accurately and had higher broad sense heritability for drought tolerance in forage grasses (Kefauver et al., 2015; Vergara-Díaz et al., 2015, 2016; Zhou et al., 2015; Gracia-Romero et al., 2017, 2018; Buchaillot et al., 2018, 2019; Fernandez-Gallego et al., 2019; De Swaef et al., 2021).

Our preliminary research suggests that RGB color space indices and derived vegetation indices may be suitable for estimation of leaf wilting, plant population, and pod yield of peanut (Balota and Oakes, 2016, 2017; Oakes and Balota, 2017). This study represents an in-depth analysis of the visual differences among healthy and wilted peanut plants and provides accurate machine learning models for estimation of peanut leaf wilting.

Materials and Methods

Results

Discussion

The 28 genotypes used in this study represent a subset of the United States mini-core peanut germplasm collection with contrasting morphological and physiological traits and with different responses to water deficit stress (Holbrook et al., 1993; Holbrook and Dong, 2005). In response to induced low soil moisture stress by covering the plots with rainout shelters from 8 to 14 WAP, the genotypes were visibly wilted with wilting scores ranging from 0 to 3 after four and from 0 to 5 after 6 weeks of stress (Figure 4). The beginning of induced moisture stress treatment coincided with beginning pegging and pod growth stages, which are growth stages sensitive to moisture stress with significant effects on yield (Rowland et al., 2007). Indeed, low soil moisture decreased peanut yield from an expected 4500 kg ha^–1 to only 2000 kg ha^–1 or less, which corresponds with peanut production in Virginia (Balota et al., 2020). For example, year 2014 was a good peanut year with a state average of 5040 kg ha^–1, but year 2010 was dry, and average yield was only 2016 kg ha^–1 (US Department of Agriculture-National Agricultural Statistics Service, 2019). This is because, under drought, chlorophyll content, carbon assimilation and photosynthetic efficiency decrease, and ultimately, biomass and yield accumulation are less (Reddy et al., 2003; Pilon et al., 2018). In this study, plots that maintained turgid vines, i.e., genotypes scored 0 and 1, had the highest yield under drought. The plots for which wilting was scored 2–5 produced only around 1000 kg ha^–1 with a 30% initial yield decrease when wilting was scored 2 followed by 50% yield reduction for plots with greater wilting (Figure 5). If the relationship between plant wilting and yield is proven on larger-scale production, a wilting score of 2 could be used to trigger irrigation of peanut, and aerial estimations could become important for efficient irrigation scheduling as suggested by Rowland et al. (2012).

The RGB-derived color indices from proximal and aerial images were associated with leaf wilting (Table 3). For example, due to water deficit, canopy color shifted from green (120°) to yellow (60°), i.e., chlorophyll content decreased, resulting in lower hue angles in wilted plants than turgid plants. Similarly, severely wilted plants had a smaller fraction of green pixels captured with canopy images and, therefore, less GA and GGA. The indices a^∗ and u^∗ increased with wilting increase; they became less negative, which indicates that canopy color changed from green to red in stressed and wilted plants. Similar results were found in soybean, wheat, corn, and peanuts confirming our hypothesis that RGB images can discriminate turgid vs. wilted peanut canopies, i.e., tissue turgor pressure falls close to zero under low soil moisture stress (Vergara-Díaz et al., 2015, 2016; Balota and Oakes, 2017; Zhou et al., 2020).

Wilting models 1 and 2 had classification accuracies of 61% for proximal and 69% for aerial images (Table 5). The accuracy improved when using the first and second highest probability for classification to 91% for proximal images and 81% for aerial images. When using the nearest score classification, wilting estimation accuracy raised to 99% for proximal images and 90% for aerial images (Table 5). Lower accuracy from aerial images could be attributed to reduced resolution in comparison with proximal images. Another reason could be the number of wilting levels, which were fewer for proximal imagery (wilting level 4 and 5 were absent) as compared with aerial images. Nonetheless, when using the nearest score classification, the accuracy of wilting estimation from aerial images increased substantially. This method is also the closest to the traditional method of visual rating. For example, when visually rating, the operator may find it difficult to assign a plot clearly to a score of 1 or 2 or rather in between 1 and 2. If rating is being performed by different operators, i.e., different visual perceptions, or at different hours of the day, i.e., different sun angles can influence visual perception, then a plot that scores 2 may be assigned 1 or 3. The wilting matrix confirmed that the majority of misclassifications were neighbors, e.g., either 2 or 4 for a score of 3. In addition, the statistical probabilities showed that the wilting scores assigned on the basis of the first highest probability were marginally ahead of the second highest probability. For example, for plot 121, the score 2 had first highest probability of 0.37, whereas the second highest probability was 0.35 for score 1 (data not shown). Therefore, using either the first or second highest probability as a true classification or nearest score classification to calibrate the process and estimate wilting scores is closer to the visual rating and had improved accuracy. The AUC values for model 1 (proximal) (0.83) and model 2 (aerial) (0.92) also support the conclusion that these models are highly predictive of peanut plant wilting. Though aerially derived wilting scores had lower classification accuracy, they correlated better (r = 0.85) and had higher predictability (AUC = 0.92) than proximally derived wilting scores (r = 0.60; AUC = 0.83) (correlation data not shown). Aerial estimation was preferred to proximal for faster data collection and analysis. It was found in this study that aerial imagery can be accomplished in a shorter period of time than proximal imaging (a 12-min flight for aerial imagery compared to 90 min for proximal imagery for 168 peanut plots), and rating using aerial sensing offers greater temporal repeatability. Although aerial imagery involves processing time (about 5 h) to orthomosaic the image, the process is fully automated at night and does not require personnel time. Moreover, manually taken proximal images were prone to errors, such as differences in camera height, sun angle, and cloud cover during the time of collection. These drawbacks were absent in aerial imagery.

Wilting models 3 (binary proximal) and 4 (binary aerial) had classification accuracies of 77% for proximal images and 88% for aerial imagery (Table 6). The AUC for model 3 was 0.82 and for model 2 was 0.87 (Figure 6). Although classification accuracy and AUC were both used for learning algorithms in machine learning, AUC is considered better and statistically more consistent than classification accuracy (Ling et al., 2003; Huang and Ling, 2005). Thus, based on AUC values, the probability of these models to estimate wilting rating was 82% when using proximal imagery and 92% when using aerial imagery. Moreover, aerially derived wilting scores correlated better with manually taken wilting scores (r = 0.71) than proximally derived wilting scores (r = 0.54) (data not shown). Aerial sensing to estimate soil moisture and trigger irrigation of crops is currently used (Ahmad et al., 2009; Ge et al., 2019). For peanut, however, with dense biomass growth and complete ground coverage early in the growing season, similar approaches may not work. Alternatively, using the binary wilting rating (turgid vs. wilted) estimated by models 3 and 4 can be used to represent plant water status. This has the potential for development of smartphone applications that can be used to decide if a peanut field requires irrigation or not, i.e., plants having turgid rating would not require irrigation, but wilted plants would require irrigation. However, this needs further investigation and feasibility analysis for implementation in peanut production.

All logistic regression models used here have high values of the mean AUC associated with cross-validation, i.e., 0.83 for model 1, 0.93 for model 2, 0.82 for model 3, and 0.93 for model 4 (Table 7). This proves the robustness of the models when used on independent data sets. Using 2019 data for validation of all models resulted in similar wilting classification accuracies as in 2018 (Tables 8, 9). This shows that the wilting estimation models developed in this study can be successfully used with other data sets.

In this article, we report new models to predict wilting in peanut and a step forward toward future automation and real-time selection for reduced wilting and improved yield and drought tolerance in the peanut breeding programs in the United States and throughout the word. We also report, for the first time, prediction models for peanut plant water status that can be useful for irrigation scheduling in the future. For example, currently, only 5% of the remote sensing models proposed to schedule irrigation are implemented at the farm level in the United States (Steve Thomason, personal communication). Our wilting approach from RGB images could be more appealing for growers to use than multispectral sensors as in other proposals, but this needs further investigation at the farm level. In a previous work, we document substantial time savings when using high-throughput techniques in breeding (Sarkar et al., 2020). At the same time, high-throughput techniques, such as reported in this article, can increase the frequency of data collection and possibly the accuracy by removing the bias of visual human perception (Milberg et al., 2008; Borra-Serrano et al., 2018; Sarkar and Jha, 2020). Overall, the logistic models developed in this study were successful in estimating peanut wilting and plant water status from RGB images, which suggests potential for aerial imagery and machine learning applications in breeding for improved drought tolerance of peanut.

Conclusion

This article reports, for the first time, on a simple yet key aspect of peanut screening for tolerance to low soil moisture stress and uses novel, fast, cost-effective, and accurate RGB-derived models to estimate leaf wilting. Leaf wilting caused by low soil moisture stress can be estimated using remote sensing. It can also be used to predict plant water status associated with maximum yield. An important aspect of this study is the combined use of remote sensing and machine learning tools to build models for high-throughput phenotyping. These methods are helpful in quick and accurate data collection in research and breeding programs and suggest good scope for further investigation of automated and efficient irrigation in peanut production.

Data Availability Statement

The software presented in this article are not readily available because of proprietary access. Requests to access the software should be directed to the corresponding author. The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author Contributions

MB wrote the grant proposal, selected the peanut genotypes to be used, and had substantial contribution at the reviewing and editing of the manuscript. A-BC helped developing protocols and routines for image collection, processing, and analysis. AR helped with the regression analysis and developing the logistic models for this study. SS did the image and data collection, their analysis, and wrote the manuscript. All authors contributed to the article and approved the submitted version.

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.