The Use of Near-Infrared Imaging (NIR) as a Fast Non-Destructive Screening Tool to Identify Drought-Tolerant Wheat Genotypes

Abstract

Due to climate change, many agricultural areas will face shortages in water availability; thus, the mission of the upcoming decades is to ensure food security while improving agriculture sustainability. The development of wheat varieties that are more adaptable to the changing climate is mandatory to achieve this goal. Genetic resources can be the key to unlock a wide genetic potential, but faster, high throughput methods are needed to easily screen the huge amount of available genetic resources. Phenotyping is the central element to exploit wheat genetic resources as it allows us to identify superior genotypes and to achieve selection gain. To select contrasting phenotypes within a core set of 149 durum wheat genotypes, belonging to the National Research Council (CNR) germplasm collection, under drought stress conditions, we studied plant water status gained by NIR imaging. By the application of the present method, it was possible to rapidly identify and select a set of putative genotypes highly tolerant to drought, as well as a set of contrasting genotypes, for further studies and/or for breeding programs. This promising approach is scalable to a larger number of genotypes in pre-breeding program.

1. Introduction

Agricultural food production is facing the huge challenge of feeding ten billion people by 2050, while reducing agricultural areas and inputs, contrasting the negative effects of climate change. This challenge can be faced only by both enhancing field yield and by increasing the sustainability of food productions.

During the 1960s, the scientific community realized the urgent necessity of preserving the declining crop genetic variation. The so-called “Plant Genetic Resources Movement” [1] began worldwide programs of collection, conservation, and utilization of crops’ germplasm.

Nowadays, it is estimated that around six million germplasm entries of different crops are conserved in gene banks worldwide. This high number of samples offers a huge wealth of gene variants and useful traits to improve plant architecture and performance improvement. Unfortunately, on the other hand, this makes the analysis of plant germplasm time and labor consuming, as well as an anti-economic enterprise, thus claiming for fast and reliable high throughput approaches to screen plant germplasm. Thanks to the developments in DNA-based technologies, genomic tools have been largely employed [2,3,4,5,6,7]. However, the identification of superior genotypes deals with the assessment of complex plant traits such as growth, development, tolerance, resistance, architecture, physiology, ecology, and yield. Refs. [8,9,10,11], describe the plant phenotype. Plant phenotype is the result of the interaction of genome and the external environmental inputs to which the plant is exposed along its entire life cycle. Therefore, the modern tendency is to associate genomic studies to high throughput phenotyping (HTP) to quickly and precisely identify plant genotypes corresponding to a specific ideotype [12,13,14] supporting and enhancing the utilization of plant germplasm. HTP technologies exploit the power of the integration of optical sensors, computer science, robotics, electronics, and artificial intelligence to extract information from plant images in a non-invasive manner, to identify digital traits associated with plant responses to stresses or to specific genotypes. HTP can be performed at different levels, from cells to plants, but also to populations, scalable to all species, including the most important food crops.

Durum wheat (Triticum durum Desf. syn. Triticum turgidum L. subsp. durum (Desf.) Husn., 2n = 4x = 28, genomic formula AABB) is a staple crop all over the Mediterranean, a fundamental source of the food tradition in this area, and a key element of the Mediterranean diet. Originally domesticated in the Mediterranean, it is nowadays grown all over the world [15]. Unfortunately, Mediterranean agriculture is strongly threatened by the increasing occurrence of drought and high temperature events due to climate change [16]. Nonetheless, the high number of durum wheat germplasm samples—the EURISCO database reports a total of over 17,000 durum accessions, some 7000 of which are traditional varieties and/or landraces—is at the same time a resource of incontestable usefulness and the major factor that hampers the possibility to explore in depth that gene pool [17,18,19,20,21].

For this reason, a program based on the Single-Seed Descent (SSD) approach was applied to obtain a core collection of durum 452 wheat genotypes, largely representative of the entire genetic variation present in the durum wheat gene pool and of the geographical origin of the material [22,23].

Drought stress is a major driver of yield losses due mostly to the ongoing climate change [24]; thus, selecting with high accuracy within large numbers of genotypes is crucial to widen the spectrum of drought resistance strategies [25].

To identify contrasting genotypes with differential responses to water stress, a high-throughput-based method centered on near-infrared (NIR) images was devised. The present communication reports on the application of this method to a core collection of durum wheat genotypes and its outcomes in terms of phenotype-driven selection.

2. Materials and Methods

2.1. Plant Material and Growth Conditions

A selection of 137 genotypes out of the core set of the 411 Single-Seed Descent (SSD) durum wheat collection were grown in a greenhouse at the National Research Council (CNR) Research Unit at Agenzia Lucana di Sviluppo e di Innovazione in Agricoltura (ALSIA)—Research Center Metapontum Agrobios (Bernalda, Italy, 40.392029° N, 16.787110° E).

The 137 genotypes were randomly selected, maintaining the high level of diversity present in the SSD collection and according to the capacity of the phenotyping platform, which cannot host more than 500 pots.

Seeds were germinated at room temperature for a maximum of 4 days on moist filter paper in Petri dishes and then transplanted into polystyrene plateau. The plateaus were then stored at 4 °C for 2 weeks to synchronize the plantlet’s growth. Individual plants were then transferred into two-liter pots filled with a 1:1 (v/v) mixture of washed river sand and peat moss until a total weight of 1200 g. Then, plants were grown in a glasshouse under natural light conditions, and environmental parameters were monitored every 30 min using a datalogger (Watchdog Model 450, Spectrum Technologies, Inc., Aurora, IL, USA).

Three replicates of each genotype were randomly distributed in the greenhouse to minimize the possible effects of microclimatic variations in the greenhouse through spatial distribution. Each single pot was identified by a barcode applied in the proper position to allow automatic reading of the plant identifier. Wheat plants were manually kept fully irrigated (100% Field Capacity, FC) up to the late tillering stage (Z30, 60 days after sowing, DAS,

Table 1. Number of reads, plant growth stage expressed in Zadoks scale.

DAS	60	67	74	81	88
Zadoks stage	31	32	33	35	36
Read number	1	2	3	4	5

). Pots were watered daily to saturation to avoid soil moisture deficit.

2.2. Imaging Experiments

Images in the visible domain (white light) for High Throughput Phenotyping (HTP) were taken every seventh day for four weeks with a Scanalyzer 3D phenotyping system (LemnaTec GmbH, Aachen, Germany), which is part of the Phen-Italy platform located at ALSIA—Metapontum Agrobios, following the procedure used by Petrozza et al., 2014 [26]. The same system has an additional imaging chamber equipped a near-infrared spectrum camera (VDS Vosskuehler NIR-300P; 900–1700 nm) whose images reflect the level of tissues water content. The system takes 3 orthogonal images (two lateral and one from above) in both visible light and NIR wavelengths. The RGB images are used to better identify gross plant morphology as well as other parameters, while the NIR shots aimed at revealing water content in leaves [27].

2.3. Data Analysis

Images were processed and analysed with the LemnaGrid software from LemnaTec GmbH. The software provides a visual programming interface where individual functions, represented by boxes, consisting of input and output ports are linked together to form a workflow for analysing images. The available processing functions offered include loading images; converting image formats; conducting image processing such as thresholding, filtering, and mathematic morphology calculations; extracting objects of interest; making measurements of said objects; and writing the measured data to a database for later statistical analysis. (https://www.lemnatec.com/image-and-data-analysis/ (accessed on 1 February 2022)).

For this work, segmentation of the NIR images consisted first of adaptive thresholding (box size 10 × 10 pixels, threshold difference of 10). Artifacts of less than 3 pixels were eliminated with the fill function by identifying a region of interest (ROI) corresponding to the area occupied by plants that was used to exclude artifacts from the imaging chamber. The intensity values of the remaining plant pixels were used for histogram analysis.

3. Results and Discussion

The approach used in this study is described in Figure 1.

In a first run, we analyzed how durum wheat plants responded to minor drought stress for NIR imaging. Images acquired by a NIR camera can be used to evaluate the plant water status based on the assumption of increased reflection in the NIR spectrum when foliar water content diminishes [27].

Typical water absorption bands in the Near-Infrared spectrum are at 1450 nm and 1930 nm; thus, the Scanalyzer NIR camera, being sensitive in the 900–1700 nm range, allows measurement of the drop in foliar reflectivity due to the 1450 nm water absorption band. An increased level of absorption will result in a reduced level of reflection in the NIR spectrum, thus allowing us to estimate leaf water content.

Both NIR and RGB images have been taken. To better define the plants’ shape and limits, the intensities of reflectance were used to quantify, in relative terms, the water content of plant tissues. These are the basis of the use of NIR imaging as a digital proxy to study water tolerance in plants [28].

NIR images are 8-bit and gray-scale, consisting of reflectance values in the range 0–255. Nevertheless, since the dynamic range of the NIR camera is fixed, under experimental conditions reflectance values occupy only a part of the whole 0–255 scale. To reduce electronic noise due to environmental conditions, the values were re-binned in the range 0 to 128 by summing the values of adjacent luminance.

To define the actual plants’ water status, the distribution of water content in the sampled plants was determined by scoring the first reading session for their pixels’ luminance together with the sum of the number of pixels belonging to each bin (0–128) recorded.

A plot of luminance against the number of occurrences was produced and analysed. The result is a distribution curve that approximates a normal distribution (Figure 2).

If the median value is considered as a discriminant, we can hypothesize that all pixels with a luminance lower than the median represent part of the plant with a good level of hydration, while all pixels with a luminance value higher than the median represent part of the plant with tissues poor in water. This is due to the response to NIR imaging, whose level of reflected luminance rises with increasing levels of drought [26,27]. On these grounds, we defined “WET pixels” those showing a luminance (L) range of 50 ≤ L ≤ 80 and representing plant parts richer in water; conversely, we named DRY pixels all pixels with a luminance value L > 80, thus representing plant parts with a lower water content (Figure 3).

In our experiment, plants were grown under regular irrigation and it was hypothesized that plants showing a higher water status can show more resilience to stress, while those with a lower hydration level can be classified as sensitive to it. Consequently, it is reasonable to assume that plants better responding to water stress should possess a higher number of “WET” pixels as compared to plants more affected by the stress.

In a second turn of analysis, all images were analyzed plant by plant, recording the number of WET pixels present in the six images taken per plant. The weighted mean of the pixels representing each plant were used to express the mean water content of that plant, so that we could order the 411 plants of the experiment based on the growing number of WET pixels recorded and listed (Figure 3,

Table 2. Mean water content recorded for each genotype (SSD Genotype) expressed as the weighted mean of wet pixels. Genotypes are ordered based on a growing number of wet pixels. se, standard error.

SSD Genotype	Mean	se	SSD Genotype	Mean	se	SSD Genotype	Mean	se
SSD393	0.1234	0.0199	SSD182	0.2577	0.0355	SSD255	0.311918	0.044896
SSD240	0.1285	0.0232	SSD336	0.2580	0.0794	SSD125	0.317003	0.031154
SSD266	0.1521	0.0309	SSD52	0.2583	0.0426	SSD412	0.31701	0.03785
SSD511	0.1550	0.0115	SSD59	0.2601	0.0176	SSD246	0.318829	0.013851
SSD322	0.1691	0.0292	SSD315	0.2627	0.0055	SSD447	0.320201	0.043648
SSD338	0.1709	0.0127	SSD220	0.2665	0.0359	SSD407	0.321094	0.037618
SSD416	0.1985	0.0141	SSD345	0.2672	0.0512	SSD467	0.323395	0.061267
SSD330	0.2016	0.0765	SSD99	0.2683	0.0549	SSD79	0.323652	0.046922
SSD397	0.2037	0.0345	SSD86	0.2706	0.0708	SSD499	0.323756	0.018017
SSD343	0.2038	0.0233	SSD96	0.2729	0.0280	SSD262	0.324868	0.062367
SSD278	0.2052	0.0355	SSD83	0.2740	0.0146	SSD281	0.326212	0.048101
SSD432	0.2052	0.0226	SSD54	0.2744	0.0211	SSD178	0.327289	0.059534
SSD128	0.2056	0.0435	SSD195	0.2746	0.0847	SSD65	0.328755	0.03194
SSD239	0.2063	0.0427	SSD415	0.2747	0.0299	SSD64	0.334353	0.056684
SSD271	0.2064	0.0676	SSD6	0.2773	0.0608	SSD244	0.335239	0.008056
SSD253	0.2083	0.0288	SSD91	0.2810	0.0395	SSD477	0.335793	0.030472
SSD269	0.2087	0.0336	SSD431	0.2817	0.0738	SSD335	0.337223	0.080187
SSD180	0.2151	0.0251	SSD424	0.2828	0.0787	SSD36	0.337427	0.026768
SSD326	0.2216	0.0304	SSD480	0.2833	0.0371	SSD70	0.338503	0.040525
SSD453	0.2237	0.0254	SSD24	0.2849	0.0507	SSD470	0.347294	0.042785
SSD155	0.2261	0.0369	SSD280	0.2858	0.0421	SSD509	0.348098	0.018469
SSD409	0.2276	0.0335	SSD288	0.2859	0.0234	SSD15	0.348783	0.050455
SSD146	0.2281	0.1248	SSD142	0.2889	0.0958	SSD231	0.351929	0.033402
SSD399	0.2306	0.0190	SSD414	0.2895	0.0456	SSD219	0.352659	0.055812
SSD112	0.2310	0.0647	SSD400	0.2896	0.0093	SSD494	0.35297	0.028618
SSD328	0.2320	0.0289	SSD348	0.2906	0.0453	SSD122	0.367246	0.00449
SSD123	0.2325	0.0651	SSD2	0.2941	0.0400	SSD256	0.36811	0.030159
SSD157	0.2329	0.0286	SSD173	0.2966	0.0542	SSD451	0.371517	0.058267
SSD43	0.2352	0.0368	SSD298	0.2977	0.0149	SSD526	0.374167	0.043784
SSD290	0.2358	0.0458	SSD7	0.2984	0.0337	SSD111	0.376019	0.011487
SSD421	0.2363	0.0210	SSD308	0.2985	0.0212	SSD227	0.378514	0.023269
SSD350	0.2370	0.0048	SSD44	0.2987	0.0525	SSD325	0.385445	0.024594
SSD487	0.2390	0.0071	SSD69	0.3000	0.0637	SSD158	0.391514	0.066688
SSD422	0.2396	0.0070	SSD426	0.3001	0.0170	SSD245	0.397842	0.064106
SSD411	0.2425	0.0340	SSD507	0.3004	0.0561	SSD162	0.400995	0.057963
SSD303	0.2450	0.0061	SSD292	0.3009	0.0239	SSD168	0.404066	0.059416
SSD237	0.2451	0.0324	SSD66	0.3010	0.0080	SSD459	0.404357	0.007627
SSD107	0.2457	0.0629	SSD137	0.3034	0.0306	SSD500	0.413246	0.078537
SSD243	0.2458	0.0623	SSD113	0.3035	0.0221	SSD171	0.415962	0.010175
SSD116	0.2484	0.0245	SSD294	0.3036	0.0441	SSD483	0.428288	0.050466
SSD441	0.2509	0.0627	SSD120	0.3037	0.0184	SSD283	0.433356	0.064512
SSD135	0.2514	0.0221	SSD423	0.3047	0.0519	SSD525	0.442475	0.011191
SSD92	0.2517	0.0157	SSD274	0.3066	0.0707	SSD505	0.450073	0.031612
SSD302	0.2545	0.0346	SSD147	0.3080	0.0631	SSD457	0.450712	0.046915
SSD443	0.2546	0.0348	SSD427	0.308078	0.060848	SSD109	0.508032	0.057536
SSD35	0.2562	0.0656	SSD513	0.311494	0.045601

).

In Figure 3, each column represents one of the 411 plants analyzed by NIR reflectance. The colored boxes represent the distribution of WET pixels in the plant—the larger the box, the higher the variation—and its height is determined by the value of the weighted mean of all readings. Thus, plants with bars in a higher position are those richer in WET pixels, and therefore plants with a better hydration level. Conversely, those with bars at lower height are those poorer in WET pixels, thus those more affected by the drought stress.

It is worth noticing that at the extreme (WET, green and DRY, red sides) plants belonging to the same genotype tend to show similar values, and thus occupy contiguous or close positions in the distribution. This is a good indication that the level of wet or dry pixels is strongly linked to the genotypic constitution of the plant and provides a simple and efficient tool to select contrasting genotypes for response to water stress.

The possibility to automate the image filtering and pixel counts suggest a simple application of this high-throughput analysis of this huge amount of data. Our method is based on the NIR reflectance of each pixel, which allows us to select different ranges of analysis to better highlight how water distributes among the plant parts, thus providing a more precise tool to analyze plants’ phenotypes.

The results of this pilot experiment demonstrate that high-throughput phenotyping tools can be successfully employed in selecting useful genotypes from germplasm collections.

The results of this experimental approach have clearly demonstrated that the application of NIR technology for high-throughput phenotyping (HTP) can be successfully employed in selecting useful genotypes from germplasm collections. NIR technology has been routinely used with hand-held instruments to analyze leaf water content in a very time-consuming procedure based on few sample points in the plant [28]. Here, the use of an NIR camera allows the rapid evaluation of plant water status in a fast and non-destructive way, allowing us to monitor one and the same plant over the time domain. This evidence reduces the bias introduced by individual variation while examining the response to an external stimulus.

Moreover, if HTP data could be associated with genomic data for single genotype, this approach will prompt the identification of genetic traits potentially associated with the desired character more rapidly and precisely in comparison with costly and time-consuming field experiments based on association approaches, such as GWAS.