Image processing and impact analyses of terminal heat stress on yield of lentil

Abstract

Abiotic factors, including heat stress, significantly impact the growth and development of lentil across the globe. Although these stresses impact the plant’s phenotypic, genotypic, metabolic, and yield development, predicting those traits in lentil is challenging. This study aimed to construct a machine learning-based yield prediction model for lentil using various yield attributes under two different sowing conditions. Twelve genotypes were planted in open-field conditions, and images were captured 45 days after sowing (DAS) and 60 DAS to make predictions for agro-morphological traits with the assessment for the influence of high-temperature stress on lentil growth. Greening techniques like Excess Green, Modified Excess Green (ME × G), and Color Index of Plant Extraction (CIVE) were used to extract 35 vegetative indices from the crop image. Random forest (RF) regression and artificial neural network (ANN) models were developed for both the normal-sown and late-sown lentils. The ME × G-CIVE method with Otsu’s thresholding provided superior performance in image segmentation, while the RF model showed the highest level of model generalization. This study demonstrated that yield per plant and number of pods per plant were the most significant attributes for early prediction of lentil production in both conditions using the RF models. After harvesting, various yield parameters of the selected genotypes were measured, showing significant reductions in most traits for the late-sown plants. Heat-tolerant genotypes like RLG-05, Kota Masoor-1, and Kota Masoor-2 depicted decreased yield and harvest index (HI) reduction than the heat-sensitive HUL-57. These findings warrant further study to correlate the data with more stress-modulating attributes.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13205-024-04031-5.

Introduction

Lentil (Lens culinaris Medik.), a leguminous plant cultivated during the winter season, is acclaimed as a rich source of essential nutrients, encompassing approximately 24–26% proteins, 57% carbohydrates, 3.2% dietary fibers, and polyphenols (Venkidasamy et al. 2019; Sonkarlay et al. 2020). Additionally, it serves as a plentiful reservoir of essential minerals and vitamins (Tiwari and Shivhare 2016; Joshi et al. 2017). These nutritional attributes place lentil as a significant source of sustenance for both human consumption and as fodder for livestock. Originating from the fertile crescent of Southeast Asia and the Mediterranean basin, lentil is presently cultivated and consumed in dry and semi-arid locations across the globe (Zohary 1972). Worldwide production of lentil was approximately 6.3 million tons from 5.5 hectares of land in 2018 (Malik et al. 2022). As the second largest producer, India contributes nearly one-fourth of the global lentil production to secure the required food demands for the expanding population (Venkidasamy et al. 2019).

Delayed sowing of lentil by deviating from the normal timeframe (usually sown within the first 2 weeks of November with temperature around 18–30 ℃) aggravates the likelihood of exposing the reproductive stages of the plants to elevated temperatures, a condition known as terminal heat stress (Choudhury et al. 2012; Sita et al. 2017; El Haddad et al. 2020). The debilitating effect of high-temperature stimuli during various reproductive phases had a negative influence on morphology and yield attributes due to disturbed phenological development in various legumes like chickpea (Kumar et al. 2015), faba bean (Abdelmula and Abuanja 2007), common bean (Seidel et al. 2016), and lentil (Sehgal et al. 2017) as well. The subsequent reason for these yield reductions can also be attributed to the repression of gene expression as well as the rapid regulation of developmental pathways and signaling (Rollins et al. 2013). As a result of exposure to high temperatures, the pathways involved in growth are severely down-regulated (Rollins et al. 2013), which can affect plant height and biomass, the two crucial yield determinants, as seen in faba bean (Siddiqui et al. 2015), chickpea (Awasthi et al. 2014) and common bean (Salazar et al. 2018). In a similar way, high-temperature stimuli have a deliberate impact on the other essential yield attributes, for instance, number of branches per plant (Aghili et al. 2012; El Haddad et al. 2020), number of filled pods and pod set (Kumar et al. 2013) as well as seed index (Sehgal et al. 2019). Furthermore, heat waves in lentil genotypes eventually reduce the harvest index (HI) by 16% (Bourgault et al. 2018).

Automated machine learning analysis based on imaging can accurately distinguish and identify plants based on their optical characteristics, shape, and texture, even when grown in complex environments such as open fields, residue, and soil ecosystems (Meyer et al. 1999). Image analysis is a mathematical technique that uses image pixels to extract, characterize, and interpret tone information, and the accessibility of information varies depending on resolution and tonal tone (Arnal Barbedo 2013). Before extracting textural features, it is necessary to identify regions of interest (ROI) or targets, which are then divided into two groups: green vegetation and non-green background (soil, rocks, weeds, and residue). The thresholding approach can be used to remove non-green background from the image, allowing for easier identification of the crop (Tang et al. 2009). Researchers have used various indices such as Color Index of Plant Extraction (CIVE) (Kataoka et al. 2003), Excess Green minus Excess Red (E × G − E × R) (Meyer and Neto 2008), and Excess Green index (E × G) in addition to threshold-based segmentation techniques to distinguish vegetation from soil (Bunting and Lucas 2006). A longer temporal duration is required for the separation of green vegetation from noise-containing color images (Kataoka et al. 2003). Applying the Modified ME × G − E × R segmentation algorithm to Multi-Scale Retinex with Color Restore (MSRCR) yielded optimal outcomes for cucumber canopy images in a naturally illuminated greenhouse (Guoxiang et al. 2016). Experimental data suggested that the Modified Excess Green Vegetation Index performs better than E × G and Otsu’s thresholding method for assessing the greenness of images while being insensitive to illuminant variations (Abdullah and Yaakob 2017). E × G − E × R was found to be more reliable than E × G + Otsu and NDVI + Otsu indices when implemented on digital color image collections of individual underdeveloped soybean plants cultivated and captured in a controlled greenhouse environment, along with fresh wheat straw backdrops (Meyer and Neto 2008). In another report, the segmentation of rice leaves impacted by leaf brown spot and blast was successfully accomplished using Vegetation Indices (VI) like Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Enhanced Vegetation Index (EVI) as well as Soil-Adjusted Vegetation Index (SAVI), and the diseased areas in the VI images were subsequently extracted using Otsu’s method (Phadikar and Goswami 2016).

The optimization of field management operations throughout the growing season can be accomplished through crop yield prediction. Such optimization requires the balanced application of fertilizers, regulation of environmental and biological stresses, suitable sowing density, and effective weed control to ensure the best possible yield in the given circumstances (Kamath et al. 2021). Although different laboratory methods have been standardized to predict crop properties, these methods are primarily destructive and time-consuming. To overcome these limitations, a non-destructive and less demanding method involving the image analysis of crops using different color values has been proposed. This method involves fitting a machine learning (ML) model with these color values and properties regarding the crop (Rasti et al. 2021). Concentrations of rice chlorophyll and leaf nitrogen have been measured using a digital still color camera together with digital color image analysis under natural light (Wang et al. 2014). Casadesús et al. (2007) highlighted the importance of comparing and establishing correlations between VIs derived from conventional digital cameras and the NDVI from spectroradiometer measurements in wheat breeding in a water-limited environment. Similarly, the study performed by Bendig et al. (2015) revealed that the integration of five visible and near-infrared VIs for biomass monitoring with drone-derived crop surface models presented a normalized ratio index (NRI)-based index that exhibited the strongest correlation (R² = 0.83) with dry biomass in barley. Following long-term agromet-spectral variables acquired from multi-sensor satellites, Prasad et al. (2021) depicted the usage of the RF algorithm for predicting cotton production in the Indian state of Maharashtra at three different stages prior to harvest. Li et al. (2020) employed the RF algorithm and sensors to obtain relative leaf area index and environmental factors, including temperature, humidity, and light intensity, to make predictions about plant transpiration. Ramos et al. (2020) used an RF ranking approach to determine which of the 33 vegetation spectral indicators derived from drone imagery were most effective at predicting maize yield. In a similar approach, Pang et al. (2022) used the RF regression technique to develop yield prediction models for three Southeast Australian wheat-growing regions. For nonlinear and challenging tasks, including agricultural output prediction, biomass change detection, and crop evapotranspiration analysis, Khan et al. (2020) applied ANN models to provide precise and realistic predictions, where the calculated biomass and field-measured biomass had an excellent association (R² = 0.76).

While numerous yield prediction models are available for various crops, limited research has been undertaken to predict yield-attributing traits in lentil under the normal- and delayed-sowing conditions. The monsoon cycle in Southeast Asia has been shifted due to erratic rainfall, which has resulted in delayed sowing and transplanting of monsoon rice (Suepa et al. 2016). Consequently, lentil sowing time has also been deferred, resulting in reduced yield due to the plant’s short life cycle and forced maturity (Mandi et al. 2015). Therefore, there is a critical need to establish a standardized algorithm that can accurately predict yield attributes in lentil for both the timely-sown and late-sown conditions. Thus, this particular study has been performed to address the following three objectives: (a) to explore the variability of lentil genotypes concerning yield-associated traits under both the normal- and delayed-sowing conditions along with the identification of the tolerant and susceptible lentil cultivars under high-temperature stimuli; (b) to predict yield attributes by analyzing RGB images of lentil plants during the normal growth stages; and (c) to identify and standardize the best ML techniques for segmenting images and predicting crop yield attributes in lentil.

Materials and methods

Genotypes

The current experimental study was undertaken using a total of 12 genotypes of lentil, which included both released varieties and advanced breeding lines., viz., HUL-57, BCL-1041, BCL-1042, Moitree, LH-84-8, DPL-62, Narendra Masoor-1, Kota Masoor-1, Kota Masoor-2, LL-699, L-4717, and RLG-05. The seeds are planted in the open field conditions of the red and lateritic zone of West Bengal to examine the impact of high temperatures and cull out lentil genotypes that are tolerant or susceptible to heat.

Planting conditions

The lentil genotypes were cultivated under open field settings on two distinct sowing dates: (i) the first week of November (6th Nov) in 2020 for the normal-sown conditions, and (ii) the first week of December (6th Dec) in 2020, for the late-sown conditions, with the intention of inducing high-temperature stress at the reproductive stages of the crop growth under the late-sown conditions. The cultivation took place at IIT Kharagpur, West Bengal, India (22° 19′ 6.58″ N and 87° 18′ 21.52″ E). During the lentil growing period, temperatures varied from 35 ºC to 22 ºC (max.) and 25 ºC to 11 ºC (min.) for the normal-sown plants and from 42 to 25 ºC (max.) and 27–14 ºC (min.) for the late-sown plants (Fig. 1). Average temperature for the normal- and late-sown condition was recorded around 21 ºC and 29 ºC for a minimum of five consecutive days, respectively, during flowering stage (Fig. 1). The genotypes were planted employing suitable plant geometry, with 3 m of row length and a spacing of 25 × 5 cm between rows and individual plants, respectively.

Fig. 1.

Temperature index (°C) showing minimum, maximum, and average temperatures during the growth season of lentil under normal- and late-sown conditions. The experiments were conducted at the departmental farm of IIT Kharagpur in the session of 2020–2021

Sample collection for yield parameter analysis

Phenotypic data were fetched from the normal-sown as well as late-sown lentil genotypes after harvesting the plants and measured for yield-attributing characteristics following earlier literature (Choukri et al. 2020). Ten plants from each genotype were chosen, and data were collected in terms of plant height, number of branches per plant, number of pods per plant, 100 seed weight, biomass, yield per plant, and HI. The value of HI was determined with the following mathematical formula,

$$\text{HI}=\frac{\text{Yield}}{\text{Biomass}}\times 100$$ 1

Statistical analysis

To analyze the data with a two-factor experimental design (temperature and genotypes), we used AGRISTAT statistical software (ICAR Research Complex, Goa, India). Mean values of the ten randomly selected plants per genotype encompassing both the normally sown and late-sown conditions were contemplated for statistical analysis. Analysis of variance (ANOVA) was conducted through the RStudio software to evaluate the variations. For each genotype, treatment, and genotype × treatment interaction, we estimated standard errors and the least significant differences (P < 0.05). The standard methodology (Singh and Chaudhary 1979) was employed to estimate phenotypic variance (σ²p), genotypic variance (σ²g), phenotypic coefficient of variation (PCV), genotypic coefficient of variation (GCV), and heritability. The linear correlation between the variables was computed using Pearson Correlation Coefficient.

Image collection for image processing

For analysis of images, pictures of ten plants from each genotype were collected at 11 a.m. with an iPhone 6 mobile. The mobile contained 8-megapixel rear iSight camera complemented by an f/2.2 aperture for improved low-light performance. It featured Focus Pixels technology for enhanced phase detection autofocus, ensuring sharp images. The camera was also supported by a True Tone flash that adjusts the flash intensity and color for more natural-looking photos. The mobile camera was placed horizontally at 32 cm higher from the soil surface level to collect the images of the lentil plants for both sowing conditions. A total of 480 images of the lentil crop were captured, from which 120 images were taken for each lentil crop’s timely and late-sowing conditions at 45 DAS and 60 DAS.

Developing rapid image segmentation method for lentil crop

In this study, Python software was used during image processing, where contrast enhancement was done by inundating the bottom and top 1% of all image pixels into the R, G, and B streams to intensify contrast and color consistency. Considering grayscale image I(x, y), in which x and y represent the row and column of pixels, the formula for calculating the contrast-enhanced image E(x, y) would be,

$${\text{Pix}}_{\left(x,y\right)}=\frac{\left({\text{Pix}}_{\left(x,y\right)}-L_0\right)\left(U_N{L}_N\right)}{\left(U_0{L}_0\right)}+L_N,L_0<{\text{Pix}}_{\left(x,y\right)}<U_0$$ 2

where U_O and U_N are the old minimum and maximum thresholds for determining the 1st and 99th percentiles of all pixel values in I(x, y), and L_O and L_N are the new minimum and maximum thresholds for determining the 0th and 255th percentiles of 8-bit images.

An assessment of vegetation’s greenness or relative density and health was provided for each image element, or pixel, as a VI. Eight VI images were created from each grayscale image once it was converted to grayscale from contrast-enhanced RGB images (Eqs. 3–10) (Du and Noguchi 2017; Yuan et al. 2019; Yuan et al. 2019; Meyer and Neto 2008; Sanjerehei 2014; Sanjerehei 2014; Du and Noguchi 2017; Sanjerehei 2014):

$$E\times G=2G-R-B$$ 3

$$ME\times G=1.262G-0.884R-0.311B$$ 4

$$CIVE=0.441RG-0.811+0.385B+18.78745$$ 5

$$E\times GR=\frac{3\times G-2.4\times R-B}{T}$$ 6

$$\text{Green blue difference}\left(\text{GBD}\right)=G-B$$ 7

$$\text{Red green difference}\left(\text{RGD}\right)=R-G$$ 8

$$\text{Green red ratio}\left(\text{GRR}\right)=\frac{G}{R}$$ 9

$$\text{Green blue ratio}\left(\text{GBR}\right)=\frac{G}{B}$$ 10

where the variables R, G, and B correspondingly represent the respective channel values for red, green, and blue in the image, and $T=R+G+B.$

Each eight VI images were then rescaled to a value between 0 and 1. The creation of the different images between ME × G and CIVE (i.e., ME × G − CIVE) was undertaken to augment the intensity contrast between pixels representing plants and background (Yuan et al. 2019). Next, utilizing Otsu’s thresholding method (Otsu 1979), a binary mask was created by following a few steps:

Let the pixels of a computer image, which is represented in L grayscale levels, be expressed as integers ranging from 0 to L − 1, and n_i represents the number of pixels in the image that have a value of i, with the dimensions $m\times n$. So, the total number of pixels will be $m\times n=n_0+n_1+...+n_{L-1}$. The normalized histogram of components, p_i was calculated with probability distribution using the formula below:

$$p_i=\frac{n_i}{\mathit{mn}}\text{where},\sum _{i=0}^{l-1}{p}_i=1,p_i\ge 0$$ 11

Then, a normalized histogram and cumulative sums, $P1(k)k\in [0,L-1]$ were computed using,

$$P\left(k\right)=\sum _{i=0}^{k}pi$$ 12

After that, cumulative means, $m(k)k\in [0,L-1]$ was measured employing the formula,

$$m\left(k\right)=\sum _{i=0}^{k}i{p}_i$$ 13

Subsequently, the global intensity mean m_G was calculated,

$$m_G=\sum _{i=0}^{l-1}i{p}_i$$ 14

Next, the between-class variance, $\sigma B2\left(k\right)k\in \left[0,L-1\right]$ was evaluated by,

$${\sigma }_B^2=\frac{{(m_G{P}_1\left(k\right)-m\left(k\right))}^2}{P_1\left(k\right)\left(1-P_1\left(k\right)\right)}$$ 15

Finally, the Otsu threshold T_otsu was obtained as the value of k for which ${\sigma }_B^2$ was the maximum.

Segmentation of the image was done using the equation:

$$\text{IS}\left(x,y\right)\left\{\begin{array}{c}1\text{if}I\left(x,y\right)>T_{\mathit{otsu}}\\ 0\text{if}I\left(x,y\right)\ge T_{\mathit{otsu}}\end{array}\right)$$ 16

where the pixel row and column are denoted by x and y, respectively.

The connected components of the image were calculated, and a field-connected component with < 2000 weight pixels was eliminated to provide a noise-free image. The final segmented images are obtained by masking noise-free threshold images.

Extraction of image features and development of machine learning models

The image was analyzed for significant information by extracting values for several VIs from different color schemes. Four common color spaces (RGB, CIE, HIS, Y’CbCr) and twenty RGB-image-based VIs were chosen to do this. A set of 35 transformed images were generated from the segmented and contrast-enhanced images listed in Table 1.

Table 1.

RGB image transformation features used for calculating the vegetation indices

Type	Name	Abbreviation	Description	References
Original	Red	R	R channel from BGR color image	Wang et al. (2014)
	Green	G	G channel from BGR color image
	Blue	B	B channel from BGR color image
Theoretical transformation	X	X	X channel from XYZ color space	Casadesús et al. (2007)
	Y	Y	Y channel from XYZ color space
	Z	Z	Z channel from XYZ color space
	L-star	L*	L channel from Lab color space	Wang et al. (2014)
	a-star	a*	A channel from CIE 1976 Lab color space
	b-star	b*	b channel from CIE 1976 Lab color space
	Hue	H	H channel from HSI color space
	Saturation	S	S channel from HSI color space
	Intensity	I	I channel from HSI color space
	Y-prime	Y’	Y’ channel from Y’CbCr color space	Liu et al. (2019)
	Cb	Cb	Cb channel from Y’CbCr color space
	Cr	Cr	Cr channel from Y’CbCr color space
Empirical transformation	Normalized red	NR	$\frac{R}{G+B+R}$	Woebbecke et al. (1995)
	Normalized green	NG	$\frac{R}{R+G+B}$
	Normalized blue	NB	$\frac{R}{G+R+B}$
	Excess red	ExR	$\frac{1.4R-G}{G+R+B}$	Meyer et al. (1999)
	Excess blue	ExB	$\frac{1.4B-G}{B+G+R}$	Guijarro et al. (2011)

The 20 experimentally transformed and 15 theoretically transferred images were used to recover four color indices that represent the entire modified image. For each transformed image, the mean (µ), skewness (θ), standard deviation (σ), and kurtosis (δ) were calculated, as reported by Kadir (2014). These were determined by adding up all the pixel values from all the modified images and presented as,

$$\mu =\frac{{\sum }_x{\sum }_{y}T\left(x,y\right)}{N}$$ 17

where T(x, y) denotes the value of a single pixel in the transformed image, and N is the total number of pixels.

The color indices for all lentil genotypes were extracted from each RGB image, resulting in a total of 140 (35 × 4) indices. These indices were then paired with corresponding target variables such as plant height, number of branches per plant, number of pods per plant, biomass, yield per plant, and HI values. ANN and RF Regression models were employed to predict these variables. For each target variable, three models were created and fitted using different sets of VIs. These sets included VIs extracted solely from 45 DAS images, VIs extracted from both 45 DAS and 60 DAS image data, and VIs extracted from only 60 DAS images. Eighty percent of the data was chosen at random for training in both regression models, while the remaining twenty percent was utilized for testing and calibrating the coefficient of determinant (R²) value for model comparison. The top models were chosen from all the deployed models, and their top 10 significant independent variables were displayed. The steps of image processing are described in Fig. 2.

Fig. 2.

Flowchart showing the process of developing a suitable machine learning technique for the prediction of yield-attributing parameters of lentil crop

Results

Evaluation of yield parameters among lentil genotypes under differential sowing dates

Details of mean performance regarding yield parameters were taken into consideration from the tested diverse lentil genotypes grown at the normal- as well as late-sowing conditions. Inspection of the plant height data revealed that LH-84-8 and DPL-62 were recorded to have maximum plant height (47 cm) in the normal-sown condition (Table 2). Still, both attributes depicted the third highest percentage (36%) of reduction in plant height at the late-sown condition. The highest percentage of reduction in plant height (52%) was shown in the HUL-57 genotype (Fig. 3a). Kota Masoor-1 was the only cultivar to be slightly higher in the late-sown condition (44 cm) than in the normal-sown condition (39 cm) (Table 2). BCL-1041 and DPL-62 had the highest number of branches (20) in the normal-sown cultivars, though DPL-62 showed the maximum percentage of reduction (77%) (Fig. 3b). HUL-57 and Kota Masoor-1 cultivars owned a greater number of branches per plant in the late-sown conditions than in the normal-sown conditions (Table 2). RLG-05 was the only genotype that had shown a larger number of pods per plant in the late-sowing (60) than the normal-sown plants (53), but the least value was calculated for this character on the first date of sowing. BCL-1041 got the maximum number of pods per plant from the normally-sown plants after harvesting, but the percentage of reduction was also huge (91%) in this genotype. Here, only four genotypes got less than a 30% reduction, and the rest of the eight genotypes showed ranges between 60%-90% reduction in the number of pods per plant (Fig. 3c). Observation regarding 100 seed weight pointed out that BCL-1041 was the only genotype to have a meager increment in seed weight under the late-sown condition. Apart from this, the % reductions of 100 seed weight were uniformly distributed between 3 and 18% in the other eleven genotypes (Fig. 3d). Among the genotypes planted in the normal planting window, L-4147 had given the highest weight of 100 seeds (3.55 gm), yet depicted maximum % reduction (18%) in second date of sowing. The highest biomass (22.35) was measured in BCL-1041 genotype along with the maximum percentage of reduction (71%) under the late-sowing conditions, while RLG-05 was counted with a minimum percentage of reduction (22%) concerning biomass (Fig. 3e). Out of 12, eight genotypes represented more than 50% reduction in their biomass in the late-sown condition. It was detected that BCL-1041 exhibited the highest yield per plant (5.63) on the first date of sowing (Table 2); even so, the genotype depicted a higher % reduction in yield per plant due to the delayed-sowing. Eight of the genotypes had received more than a 70% reduction, where HUL-57 happened to be the most reduced on the second date of sowing, considering a decrease in yield. RLG-05 had experienced a minimum decline (40%) in yield per plant % (Fig. 3f). Analysis depicted that Moitree was the cultivar with the highest HI (36.64) under the normal date of planting (Table 2), unlike the previous results, percentage reduction of this trait was also generous (36%) among the cultivars in the delayed-sowing condition. In contrast, Narendra Masoor-1 had acquired the maximum drop (58%) in the % of HI in the delayed-sowing but had the second highest HI in the normal-sown circumstances (Fig. 3g). The maximum critical difference (CD) values at significance levels of 1% and 5% were determined for the number of pods per plant, with plant height following closely behind. The late-sown plants had higher CD values for only these two parameters compared to the normal-sown ones. On the contrary, the late-sown lentils were assessed with lower values across the board for the remaining five characteristics, i.e., number of branches per plant, biomass, 100 seed weight, yield per plant, and HI, compared to the normal-sown one (Table 2).

Table 2.

Phenological, morphological and yield traits of 12 lentil genotypes under the normal-sown and late-sown (LS) conditions with the value of mean ± SD

Sl. No	Genotypes	Plant Height(cm)		Number of branches per plant		Number of pods per plant		100 seed weight (g)		Biomass (g)		Yield per plant (g)		Harvest index
		Normal-sown	Late-sown	Normal-sown	Late-sown	Normal-sown	Late-sown	Normal-sown	Late-sown	Normal-sown	Late-sown	Normal-sown	Late-sown	Normal-sown	Late-sown
1	HUL-57	40 ± 2.53	19 ± 1.47	5 ± 1.46	6 ± 1.06	192 ± 47.57	18 ± 2.53	2.35 ± 0.38	2.16 ± 0.04	8.84 ± 2.23	3.04 ± 1.80	2.73 ± 0.45	0.52 ± 0.07	30.88 ± 3.32	17.06 ± 4.05
2	BCL-1041	45 ± 3.83	41 ± 1.47	20 ± 2.20	13 ± 1.58	361 ± 27.35	34 ± 4.11	2.62 ± 0.07	2.64 ± 0.06	22.35 ± 1.93	6.42 ± 1.04	5.63 ± 0.73	1.26 ± 0.33	25.18 ± 3.55	19.62 ± 3.95
3	BCL-1042	40 ± 1.99	34 ± 2.59	16 ± 2.91	4 ± 0.94	168 ± 28.59	43 ± 6.19	3.23 ± 0.05	2.72 ± 0.08	13.57 ± 2.42	5.14 ± 1.65	3.51 ± 0.88	1.00 ± 0.19	25.85 ± 3.92	19.53 ± 4.64
4	Moitree	28 ± 2.80	23 ± 1.83	9 ± 1.88	5 ± 1.13	132 ± 41.35	47 ± 2.69	2.30 ± 0.41	2.23 ± 0.05	9.29 ± 2.88	4.40 ± 0.57	3.41 ± 0.49	1.04 ± 0.21	36.68 ± 4.37	23.63 ± 4.59
5	LH-84–8	47 ± 2.35	30 ± 1.88	16 ± 1.73	7 ± 0.88	170 ± 13.40	37 ± 1.59	2.56 ± 0.04	2.30 ± 0.08	10.56 ± 1.55	6.06 ± 1.54	2.70 ± 0.43	0.79 ± 0.06	25.57 ± 4.54	13.04 ± 4.42
6	DPL-62	47 ± 1.97	30 ± 2.98	20 ± 1.58	5 ± 1.25	193 ± 22.02	41 ± 4.17	2.30 ± 0.03	2.20 ± 0.17	20.36 ± 1.35	7.21 ± 0.51	3.36 ± 0.69	0.71 ± 0.10	16.52 ± 3.48	9.80 ± 3.84
7	Narendra Masoor-1	37 ± 2.03	20 ± 1.79	8 ± 1.11	5 ± 1.03	70 ± 9.09	59 ± 1.99	2.52 ± 0.03	2.41 ± 0.03	4.83 ± 0.98	3.12 ± 0.53	1.70 ± 0.20	0.46 ± 0.10	35.12 ± 3.86	14.67 ± 3.54
8	Kota Masoor-1	39 ± 5.25	44 ± 3.34	11 ± 2.09	15 ± 1.30	113 ± 22.36	81 ± 8.31	3.20 ± 0.12	2.80 ± 0.09	17.05 ± 1.37	8.58 ± 0.72	2.92 ± 0.54	1.32 ± 0.22	17.12 ± 3.68	15.43 ± 4.11
9	Kota Masoor-2	27 ± 1.68	27 ± 1.54	5 ± 1.64	4 ± 1.19	69 ± 11.62	52 ± 6.22	2.80 ± 0.72	2.40 ± 0.10	5.59 ± 1.55	3.35 ± 0.72	1.80 ± 0.23	0.90 ± 0.16	32.25 ± 4.05	26.74 ± 3.87
10	LL-699	34 ± 2.56	25 ± 1.71	6 ± 1.55	5 ± 1.16	121 ± 24.86	29 ± 2.29	2.60 ± 0.11	2.30 ± 0.08	9.03 ± 1.75	3.42 ± 0.76	2.20 ± 0.24	0.55 ± 0.06	24.36 ± 4.43	16.05 ± 3.45
11	L-4717	31 ± 1.87	25 ± 1.86	7 ± 1.01	4 ± 1.20	103 ± 14.80	32 ± 2.02	3.55 ± 0.03	2.90 ± 0.69	8.65 ± 2.42	4.10 ± 0.79	2.60 ± 0.41	0.63 ± 0.10	30.02 ± 5.34	15.42 ± 3.21
12	RLG-05	43 ± 3.37	35 ± 1.64	8 ± 1.46	8 ± 1.16	53 ± 15.73	60 ± 17.95	3.20 ± 0.40	2.90 ± 0.12	7.34 ± 2.83	5.75 ± 0.68	1.70 ± 0.26	1.02 ± 0.21	23.20 ± 3.89	17.80 ± 2.82
	Critical Difference (CD) 5%	9.54	9.82	7.37	3.84	10.88	14.13	0.36	0.29	7.52	1.79	1.12	0.34	7.59	5.21
	Critical Difference (CD) 1%	13.46	13.85	10.41	5.42	15.36	19.94	0.51	0.41	10.61	2.53	1.58	0.48	10.71	7.36

Fig. 3.

Graphical representations of percentage reduction for a plant height, b number of branches per plant, c number of pods per plant, d 100 seed weight, e biomass, f yield, and g harvest index under late sowing condition compared to the early sowing condition. The value represents mean ± SD

Genetic variability appraisal of lentil genotypes under differential sowing dates

The analysis of variance revealed significant variations among 12 lentil cultivars in relation to seven yield attributes, thereby confirming the existence of substantial genetic variability (Table 3). In both sowing conditions, it was observed that the PCV% exhibited a significantly greater magnitude compared to the GCV%, which indicated an environmental effect for the majority of the yield-attributing characteristics (Table 3). GCV% and PCV% were found to be the highest for the average number of branches per plant during both the sowing windows. Least GCV% and PCV% were detected in plant height and 100 seed weight for the normal-sown and the late-sown conditions, respectively. The heritability in a broad sense across all the seven yield attributes was high and varied between 68.01 and 88.73% in the normal-sown condition and 64.57–86.63% in the late-sown state. For both the sowing conditions, maximum and minimum heritability were determined for 100 seed weight and plant height, respectively. Genetic advance as percentage of mean (GAM) varied from 27.67 for plant height to 79.32 for number of branches per plant in the normal-sowing environment. The GAM exhibited a range of values from 25.13 to 79.32, as observed over the two planting dates reported in this study, indicating that the genetic progress of the genotypes was relatively high (> 20%) for all the evaluated variables.

Table 3.

Variability, heritability and expected genetic advance of relevant yield parameters of the (a) normal-sown and (b) late-sown lentil plants

Trait	Normal					Late
	Genotypic coefficient of variation (%)	Phenotypic coefficient of variation (%)	Heritability in Broad sense (%)	Genetic advance	Genetic advance as percentage of mean	Genotypic coefficient of variation (%)	Phenotypic coefficient of variation (%)	Heritability in broad sense (%)	Genetic advance	Genetic Advance as percentage of mean
Plant height (cm)	16.29	19.75	68.01	10.73	27.67	17.11	21.29	64.57	9.97	28.32
Number of branches per plant	46.06	55.1	69.89	8.79	79.32	35.45	43.59	66.13	4.08	59.39
Number of pods per plant	34.59	36.75	88.59	26.72	67.07	34.28	37.24	84.73	28.68	65
100 seed weight (g)	16.31	17.31	88.73	0.9	31.64	13.11	14.08	86.63	0.64	25.13
Biomass (g)	43.2	51.85	69.42	8.83	74.14	29.26	35.36	68.48	2.05	49.88
Yield per plant (g)	36.97	41.01	81.3	1.97	68.67	31.04	35.74	75.45	0.48	55.55
Harvest index	22.56	25.87	76.03	11.04	40.52	22.07	25.8	73.17	6.89	38.88

Correlation study of lentil genotypes under differential sowing dates

The correlation coefficients of yield per plant with the remaining six investigated variables in both the normal-sown and late-sown experiments are presented in Table 4. The results obtained in this study revealed a strong and statistically significant positive correlation between the yield per plant and the number of pods per plant (r = 0.93**), as well as biomass (r = 0.83**) under the normal-sown conditions. Yield per plant also exhibited positive correlations with morphological traits such as number of branches per plant (r = 0.72**) and plant height (r = 0.36*). A negative non-significant correlation was detected between 100 seed weight (r =  − 0.16) and yield per plant in the presence of the normal-sown environment (Table 4). Similarly, yield per plant showed highly significant correlations with plant height (r = 0.85**), number of branches per plant (r = 0.70**), and biomass (r = 0.69**) upon the late-sown condition. Interestingly, low yet positive and meaningful correlation was perceived between yield per plant alongside 100 seed weight (r = 0.48**) and HI (r = 0.38*) under the late-sown conditions (Table 4).

Table 4.

Correlation coefficients between different trait combinations based on 12 lentil genotypes under the (a) normal-sown and (b) late-sown conditions

Trait	Normal							Late
	Yield per plant (g)	Plant Height (cm)	Number of branches per plant	Biomass (g)	Number of pods per plant	100 seed weight (g)	Harvest index	Yield per plant (g)	Plant Height (cm)	Number of branches per plant	Biomass (g)	Number of pods per plant	100 seed weight (g)	Harvest index
Yield per plant (g)	1.00							1.00
Plant height (cm)	0.36*	1.00						0.85**	1.00
Number of branches per plant	0.72**	0.71	1.00					0.70**	0.80	1.00
Biomass (g)	0.83**	0.59	0.82	1.00				0.69**	0.85	0.68	1.00
Number of pods per plant	0.93**	0.52	0.69	0.78	1.00			0.52**	0.49	0.41	0.48	1.00
100 seed weight (g)	− 0.16	− 0.13	− 0.17	− 0.08	− 0.33	1.00		0.48**	0.61	0.38	0.34	0.43	1.00
Harvest index	− 0.21	− 0.66	− 0.51	− 0.68	− 0.23	− 0.21	1.00	0.38*	− 0.05	− 0.07	− 0.38	0.08	0.05	1.00

*, **Represent the Significance at p < 0.05 and 0.01, respectively. The significance tests were performed with the correlation coefficients of yield per plant with the remaining six investigated variables only

Rapid image segmentation

Although contrast enhancement (Fig. 4) was applied in this study, it had little impact on the output images, as the mobile phone camera had already captured images with adequate natural illumination and moderate contrast during image collection at 11 a.m. Three vegetative indices, namely CIVE, ME × G − CIVE, and RGD, were found to be highly effective in separating the green background, resulting in almost black backgrounds with minimal background noise (Fig. 5). In addition, when using Otsu’s threshold approach for thresholding (Fig. 6), vegetation indices such as CIVE, ME × G − CIVE, RGD, and GRR provided excellent background separation from the green portion of vegetation images with near-black backgrounds and minimal background noise. The remaining noise was removed by utilizing connected components, and ME × G − CIVE performed exceptionally well in generating masks for further use (Fig. 7; Supplementary material 1).

Fig. 4.

Contrast enhanced images of the normal-sown HUL-57 and BCL-1041 lentil genotypes at 45 DAS, a original (HUL-57), b contrast enhanced (HUL-57), c original (BCL-1041), and d contrast enhanced (BCL-1041)

Fig. 5.

Vegetation index images of the normal-sown HUL-57 lentil genotype at 45 DAS, a original (HUL-57), b E × G (HUL-57), c ME × G (HUL-57), d CIVE (HUL-57), e ME × G − CIVE (HUL-57), f E × GR (HUL-57), g GBD (HUL-57), h RGD (HUL-57), i GRR (HUL-57), and j GBR (HUL-57)

Fig. 6.

Threshold images of the normal-sown HUL-57 lentil genotype at 45 DAS, a original (HUL-57), b E × G (HUL-57), c ME × G (HUL-57), d CIVE (HUL-57), e ME × G − CIVE (HUL-57), f E × GR (HUL-57), g GBD (HUL-57), h RGD (HUL-57), i GRR (HUL-57), j and GBR (HUL-57)

Fig. 7.

Masked images of the normal-sown HUL-57 and BCL-1041 lentil genotypes at 45 DAS, a original (HUL-57), b masked (HUL-57), c original (BCL-1041), and d masked (BCL-1041)

Extraction of image features and comparison between prediction models

The yield per plant displayed the highest degree of positive skewness among the six studied dependent variables (Table 5). Additionally, plant height and the number of pods per plant showed moderately low skewness values of 1.72 and 1.71, respectively. For the normal-sown lentil images, three separate ANN models were fitted for independent target variables, using VIs extracted from the 45 DAS, 60 DAS, and both of the 45 DAS as well as 60 DAS image data (Table 6). The yield per plant model of the ANN model had the highest R² value (0.48) with the 45 DAS image data for the normal-sown condition, while the outcomes of this model did not meet the expected standards. Similar ANN models were developed for the late-sown lentil images at 45 DAS, 60 DAS, and both 45 DAS and 60 DAS. In this condition, the model using the 45 DAS image data produced better results compared to the other models (Table 6).

Table 5.

Summary statistics of target variables

	Plant height (cm)	Number of branches per plant	Number of pods per plant	Biomass (g)	Yield per plant (g)	Harvest index
Mean	37.80	10.96	144.66	11.40	2.85	26.89
Standard deviation	7.35	5.61	83.36	5.78	1.17	7.23
Min	23.00	3.00	25.00	3.19	1.25	9.33
Max	51.00	23.00	407.00	26.32	6.76	43.18
Skewness	1.72	3.08	1.71	3.03	3.30	2.33

Table 6.

Results of R² and MAPE values for ANN and RF model with 45 DAS, 60 DAS and combining 45 DAS and 60 DAS using the normal- and late-sown lentil image data

Target variable	Normal									Late
	ANN			RF						ANN			RF
	45 DAS	60 DAS	45 DAS + 60 DAS	45 DAS		60 DAS		45 DAS + 60 DAS		45 DAS	60 DAS	45 DAS + 60 DAS	45 DAS		60 DAS		45 DAS + 60 DAS
	R2	R2	R2	R2	MAPE	R2	MAPE	R2	MAPE	R2	R2	R2	R2	MAPE	R2	MAPE	R2	MAPE
Plant height (cm)	0.15	0.12	0.00	0.75	7.64	0.67	8.66	0.66	10.74	0.45	0.00	0.24	0.58	0.89	0.42	16.22	0.53	12.76
Number of branches per plant	0.12	0.10	0.05	0.83	18.17	0.75	34.79	0.68	36.25	0.56	0.09	0.16	0.56	0.9	0.53	36.25	0.56	40
Number of pods per Plant	0.42	0.01	0.00	0.81	17.91	0.87	27.89	0.83	15.76	0.42	0.03	0.22	0.69	0.94	0.85	17.73	0.70	20.6
Biomass (g)	0.41	0.02	0.00	0.54	39.31	0.44	35.82	0.42	22	0.48	0.00	0.10	0.54	0.91	0.52	38.663	0.61	28.84
Yield per plant (g)	0.48	0.00	0.19	0.72	23.00	0.69	19.16	0.74	17.42	0.56	0.11	0.44	0.79	0.92	0.53	24.44	0.80	13.6
Harvest index	0.05	0.00	0.00	0.22	19.08	0.33	19.08	0.35	23.92	0.27	0.00	0.25	0.61	0.87	0.29	43.02	0.59	29

ANN artificial neural network, RF random forest, DAS days after sowing, R² coefficient of determinant, MAPE mean absolute percentage error

In addition to the ANN models, three different RF regression models were fitted for the normal-sown lentil genotypes using VIs extracted from the 45 DAS images, 60 DAS, and both 45 DAS as well as 60 DAS image data (Table 6). The yield per plant and plant height exhibited satisfactory findings in both 45 DAS and 60 DAS photos, accompanied by better results from the 45 DAS crop images on the first date of sowing. From the 45 DAS lentil crop images, the number of branches per plant displayed the maximum R² value (0.83), whereas the number of pods per plant depicted a slightly lower value (0.81) in this condition. Based on the data collected from 60 DAS images, the observation of the number of pods per plant exhibited the highest R² value of 0.87. Conversely, the plant height data from the normal-sown plants demonstrated relatively low mean absolute percentage errors (MAPE) of 7.64% and 8.66% in the 45 DAS and 60 DAS, respectively. HI had a minimum R² value from 45 DAS (0.22) and 60 DAS (0.33) lentil images in the same environment. Models fitted from 45 DAS crop data for the normal-sown plant images delivered better results than 60 DAS except for the number of pods per plant and HI (Table 6).

Similarly, images from the late-sown lentil genotypes at 45 DAS, 60 DAS, and both 45 DAS as well as 60 DAS (Table 6), were selected as random variables for fitting into separate RF regression models. The yield per plant demonstrated a higher R² value of 0.79 for 45 DAS crop image data, and the number of pods per plant displayed R² value of 0.85 for the 60 DAS image data in the late-sown condition. HI bestowed a good R² value from 45 DAS image data (0.61) compared to 60 DAS (0.29). The RF models outperformed the ANN models in all evaluated models. For the number of branches in the case of 45 DAS normal-sown data, there was a massive difference between the R² values of the RF model (0.83) and the ANN model (0.12). Plant height was one of the target variables that showed a massive difference in prediction between these models in every criterion. The RF model for the late-sown 45 DAS lentil models successfully predicted all the target variables, and the accuracy scores for the number of branches per plant from both models were identical (R² = 0.56) (Table 6). Best-performing models for each target variable with the normal- and late-sown lentil data images were enlisted in Table 7.

Table 7.

List of best performing models for the normal- and late-sown lentil data images

Target variable	Independent variables
	Normal			Late
	45 DAS	60 DAS	Both 45 DAS and 60 DAS	45 DAS	60 DAS	Both 45 DAS and 60 DAS
Plant height (cm)	✓			✓
Number of branches per plant	✓			✓
Number of pods per plant		✓			✓
Biomass (g)	✓					✓
Yield per plant (g)			✓			✓
Harvest index			✓	✓

DAS days after sowing

Discussion

Due to global warming and variable weather conditions, the lentil crop encounters elevated temperature stress at various growth stages. This stress can significantly affect its economic significance related to the timing and intensity of the stress (Yadav et al. 2007). Lentil has enormous susceptibility to high-temperature stimuli, particularly during the crucial periods for flowering and grain-filling. Temperatures over 32 ºC have the potential to impede numerous physiological processes within plants, including the rate of photosynthesis and respiration, metabolism, and electron transport (Redden et al. 2014), which eventually causes a lesser quantity of pod formation, occurrence of infertile pollen, and abortion of flower in lentil (Bhandari et al. 2016; Sita et al. 2017). Grain yields suffer greatly as a consequence of these alterations (Sita et al. 2017; Kumar et al. 2016). The optimal temperature for the maturation stage of lentil plants is situated within the range of 18–30 °C, which is higher than the lower temperatures required during the vegetative stage of growth (Sehgal et al. 2017). The majority of lentil genotypes identified in this study under the late-sowing conditions depicted moderate to drastic reduction in various yield-attributing traits (Fig. 3), which might be due to the average day temperature of more than 25 °C during the crop flowering stage (Fig. 1). Heat stress also distorts the water relations, respiration, primary and secondary metabolites production, along with hormone synthesis (Redden et al. 2014). According to earlier literature, exposure to high temperatures before flowering also damages the morphology and the number of flowers as well as fruits, along with the decrease in yield for this crop (Bhandari et al. 2016). The deliberate practice of planting crops later in warmer temperatures is a commonly used method to induce heat stress during the flowering and pod-filling stages, referred to as terminal heat stress, also examined in the context of this study. Results accumulated from this report suggested the adverse impact of high-temperature stress in the late-sown lentil plants during the flowering and seed-filling stage along with phenotypical deformation including growth, yield components, and shortened duration of flowering to ensure force transit from vegetative to the reproductive stage in comparison to the normal-sown plants, which has also been discussed in preliminary investigations on lentil (Sehgal et al. 2017), common bean (Vargas et al. 2021) and chickpea (Awasthi et al. 2014). Similarly, quite a few researches on members of the Fabaceae family of plants, including chickpea (Awasthi et al. 2014), common bean (Kazai et al. 2019), and faba bean (Abdelmula and Abuanja 2007) revealed a striking reduction in the plant height and total biomass, which agreed upon current findings and was associated with inhibiting the metabolism of growth-related expression (Rollins et al. 2013). In the late-sown lentil plants, number of branches formation and 100 seed weight were reduced in the introduction to the high-temperature stimuli, and these factors have significant correlations with the yield as evidenced from prior analyses on lentil (Aghili et al. 2012; Kumari and Chandra 2011, El Haddad et al. 2020) since higher yield is produced from plants with production of more branches (Mondal et al. 2013). The present investigation observed a reduction in harvest index (HI) ranging from approximately 30% to 60% among the various lentil genotypes under heat stress conditions. This reduction is notably higher compared to the findings reported by Bourgault et al. (2018). Also, HI exhibited a positive correlation with yield, as previously demonstrated in studies focusing on lentil (Rahimi et al. 2016) and chickpea (Krishnamurthy et al. 2013). Consistent with prior research, the present investigation found that the number of pods per plant declined in response to elevated temperatures, perhaps because heatwaves in the duration of the pod-filling stages trigger flower and pod abortion in lentil (Bhandari et al. 2016). Loss of pollen germination and viability due to fewer number of pods in chickpea (Devasirvatham et al. 2012) and soybean (Djanaguiraman et al. 2013) might also contribute to a drop in yield. Therefore, the study on lentil genotype (Choukri et al. 2020), which is in agreement with this report, clearly observed the positive correlation between yield and number of pods per plant. Additionally, these two significant attributes, i.e., yield per plant and number of pods per plant were found to be the most suitable plant factors for early prediction of lentil production in both conditions through Random Forest regression models developed in this study. The genotypes RLG-05, Kota Masoor-1, and Kota Masoor-2 were determined to be relatively tolerant to high-temperature stimulation after evaluation of all yield metrics and assessment of the image analysis factors. All the aforementioned elements were taken into consideration, and it was detected that HUL-57 was mostly sensitive under the late-sowing environment as well as terminal heat stress. In addition to this genotype, analysis of the majority of the yield metrics revealed that BCL-1041, LH-84-8, DPL-62, and BCL-1042 had demonstrated heat susceptibility.

This study also demonstrated that significant genotypic variability existed in relation to yield-attributing traits through a combination of environmental factors and different genetic components of lentil cultivars. Similar to previous reports (Vanave et al. 2019; Veni et al. 2020), it was depicted that PCV% gave higher values than GCV% in this study. Furthermore, excluding 100 seed weight, the differences between GCV% and PCV% were substantially higher, suggesting a distinct role of environmental stimuli for phenotypic expression of the mentioned characteristics together with the genotypic effect, which is in accordance with Sharma et al. (2010). For the evaluation of heritability, the span of 60–79% was moderately high, and more than 80% was very high (Singh 2001). Certain characteristics, including plant height, number of branches per plant, and biomass, did not show a very high heritability, suggesting that environmental factors might play some role in determining phenotypic expression (Yanti 2016). Similar findings were also perceived in the late-sown plants, which had a lower heritability amount than the normal-sown plants. Since heritability alone cannot measure genetic improvement, a combination of medium to high GA along with heritability indicated additive gene action in the inheritance of genetic traits (Ogunniyan and Olakojo 2014; Ghosh et al. 2022). Finally, a reduction in the amount of heritability, GA, and GAM, as well as overall genetic variability of the yield-attributing parameters, specified the effect of environmental aspects in the presence of the late-sown condition, possibly due to terminal heat stress.

The correlation between phenotype and genotype is dependent on the surrounding environment, whereas plant phenotypes may change at various stages of growth and affect plant performance, but the genotype of plants remains constant (Chen et al. 2014). Therefore, early prediction of plant attributes can be defined as the prediction of yield attributes, such as yield per plant, based on plant phenotyping at an early season to speed up the plant selection process for crop breeding because breeders need not wait for a whole crop production cycle (Elibox 2012). Numerous research papers explained the significant impact of early prediction of plant traits based on their phenotypical performance. One of the examples is the prediction of soybean yield by analyzing the canopy reflectance of the plant at various stages of reproduction using NDVI (Ma et al. 2001). Similarly, three plant traits, i.e., yield, seed size, and maturity, were early predictable attributes for the same crop through canopy RGB values (Yuan et al. 2019). Notably, the RGB model was also used to forecast grapevine yield-based on the quantity of berries produced throughout the growth phase of the fruits (Aquino et al. 2018). In another report, estimation of sugarcane juice Brix and fiber content before three months of the harvest provided an accurate prediction of those values at the maturity stage (Elibox 2012). The current investigation aims to identify a suitable model for the prediction of yield-attributing traits of lentil sown at two different dates considering the images captured at mid-vegetative (45 DAS) and pre-flowering stages (60 DAS) of crop growth.

The images captured under natural lighting were analyzed using various vegetation indices, including CIVE, ME × G-CIVE, and RGD. The study found that these three indices produced the best results in separating the green parts from the background with minimal noise. Furthermore, when combined with the GRR indices, these three indices provided even better results when using Otsu’s threshold method. ME × G-CIVE with Otsu’s thresholding was found to be the best method for lentil crop segmentation in both the normal- and late-sowing environments. In prior research conducted by Guoxiang et al. (2016), the modified ME × G-E × R segmentation algorithm was found to generate the most reliable results for canopy images of cucumber in a greenhouse with a variety of natural illumination settings. Additionally, Abdullah and Yaakob (2017) reported that the Modified Excess Green Vegetation Index outperformed the traditional E × G approach with Otsu’s thresholding in determining the greenness of an image, as it was less sensitive to illuminant fluctuations. Incorporating local density estimates using vegetation indices can increase the accuracy of segmentation techniques and make it easier to analyze images of various sizes captured with low-cost devices in multiple environments.

Vegetative indices were derived from lentil crop images taken at 45 DAS and 60 DAS under both the normal-sown and late-sown conditions to calculate the respective mean, standard deviation, skewness, and kurtosis values (Table 5). A total of 140 indices were developed from combinations with R, G, and B estimations to create a particular model regarding the early prediction of lentil yield attributes. Prior research has examined the potential of implementing color and texture features from canopy RGB images to forecast soybean traits with the most successful outcomes achieved by applying all 457 predictor variables in combination with the Cubist regression method and the RF classification algorithm (Yuan et al. 2019). Using an elliptical model fitting in the Cr − Cb color space after converting from the RGB color space to the Y’CbCr color space, 93.5% accuracy was attained in distinguishing between immature and mature pomelo fruits (Liu et al. 2019). RF was found to be a better model than the ANN for all the images considered in the current study. Similarly, the RF and support vector regression (SVR) models were found to be effective in predicting crop height in corn, with the former providing higher overall accuracy than the latter (Xie et al. 2021). The present report demonstrates that the RF model for number of pods per plant revealed strong performance, achieving an excellent accuracy score (R² = 0.87) for the normal-sown lentil genotypes at 60 DAS (Table 6). Additionally, the RF model for yield per plant in the late-sown lentil with both 45 DAS and 60 DAS images demonstrated a high level of accuracy (R² = 0.80) (Table 6). The study also found that crop genotype was the most vital factor for the normal-sown lentil, with a variable importance of around 66.72% (Supplementary Material 2). Another successful example of early prediction through machine learning methods was observed in the construction of an efficient method for potato crop emergence rate and symmetry using RGB ortho-images acquired by a drone along with Excess Green Index and Otsu thresholding methods, which were analyzed by an RF classifier that provided an excellent R² of 0.96 compared to manual evaluation (Li et al. 2019).

Overall, the combined application of image processing alongside machine learning approaches has provided a viable solution to the challenge of early and accurate identification of any kind of stress in plants. Disease detection and recognition in the preliminary phase have been accomplished using image segmentation algorithms (Singh et al. 2015). Image pre-processing has also been utilized to diagnose diseases like cucumber powdery mildew, speckle, and downy mildew in plants (Ying et al. 2008). In addition, the type of pest inciting the disease has also been analyzed with the help of an autonomous system with computer vision techniques (Gondal and Khan 2015). The use of multivariate analysis in the current study has offered a valuable tool for identifying the factors that influence the variation in plants. Identification of phenotypic expression has been made possible by Green Red Vegetative Indices (GRVI) to detect the seasonal variation of plants (Motohka et al. 2010). The current study has found that yield attributes such as yield per plant and number of pods per plant can be early predicted with image segmentation and ML algorithms using the images clicked by a smartphone. This finding highlights the potential benefits of developing mobile apps with suitable image analysis and machine learning algorithms for early prediction of yield, stress detection, and disease diagnosis in crops, providing significant value to crop breeders and farmers.

Conclusions

This study explored an early prediction of the yield performance of lentil crop with non-destructive and automated techniques based on RGB images captured by a smartphone. From developed models, ME × G–CIVE with Otsu’s threshold was the best in image segmentation, and RF regression was a better ML algorithm among the compared techniques to predict yield parameters for lentil. Comparing the normal-sown and late-sown image data, yield per plant and number of pods per plant were probable traits that might be predicted with RGB imagery. Further studies could enhance the disclosure of more precise outcomes for each lentil genotype by developing prediction models that are specific to the phenological stage of the crop. The result of the study also represented that the rise in temperatures caused a negative impact on yield and comprehensive agro-morphological parameters in lentil. Alarming reductions in yield per plant, biomass, plant height, number of branches per plant, number of pods per plant, HI, and overall plant health were detected in warmer average day temperatures than the optimum temperature during reproductive phases like flowering and pod-filling. In-depth research will be required to completely comprehend the mechanism underlying the heat-related impact on the quality and quantity of each evaluated lentil genotype. Due to global warming, lentil is expected to encounter recurrent and intense high-temperature stimuli in rainfed and arid regions. To ensure the production of lentils in a sustainable manner, it is crucial to identify terminal heat-tolerant cultivars with high yields and higher economic importance through the utilization of conventional breeding as well as suitable image processing techniques-based phenotyping.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Hena Gain: data curation, investigation, methodology, formal analysis, validation, visualization, and writing—original draft. Ruturaj Nivas Patil: investigation, data curation, formal analysis, writing—original draft, validation, and software. Konduri Malik: investigation, data curation, formal analysis, writing—original draft, and software. Arpita Das: supervision and writing—review. Somsubhra Chakraborty: conceptualization, investigation, methodology, supervision, writing—review and editing. Joydeep Banerjee: conceptualization, investigation, methodology, supervision, writing—review and editing.

Funding

The funding for this publication was provided by the Sponsored Research and Industrial Consultancy, Indian Institute of Technology Kharagpur.

Declarations

Conflict of interest

The authors declare that they have NO discernible competing financial interests or personal affiliations that could have conceivably impacted the conclusions provided in this manuscript.

Research involving human participants and/or animals

This work does not feature any types of animal or human participants, so the specific section is not applicable.

Informed consent

The authors have revised and approved the submitted work, confirming that no part of it has been published or is under consideration for publication elsewhere. Each author has made substantial contributions to the text and takes full responsibility for its content, format, and accuracy.