Machine learning prediction of pre-harvest fruit quality

Abstract

A computer-implemented method comprising: receiving, as input, spectral image data acquired with respect to a unit of tomato; calculating, from the spectral image data, spectral reflectance values associated with the unit of tomato in at least one specified combination of spectral bands; deriving, from the spectral reflectance values, a normalized difference spectral index (NDSI) value; and determining, based on the NDSI value, a value of at least one specified quality parameter associated with the unit of tomato.

Description

MACHINE LEARNING PREDICTION OF PRE-HARVEST FRUIT QUALITY CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority from U.S. Application Ser. No. 63 / 719,994, filed November 13, 2024, entitled “MACHINE LEARNING PREDICTION OF PREHARVEST FRUIT QUALITY,” the contents of which are hereby incorporated herein in their entirety by reference.FIELD OF THE INVENTION

[0002] The present invention relates to the field of machine learning.BACKGROUND

[0003] Tomato (Solarium Lycopersicori) is one of the most substantial crops grown worldwide, with significant economic and nutritional values. The global gross production of tomatoes was 189 million tons in 2020, valued at US $100 billion. Beyond the economic value, tomatoes are nutritionally rich, offering sugars, organic acids, lycopene content, and ascorbic acid (vitamin C), which are substantial requirements for the human diet. The consumption of tomatoes has been associated with risk reduction in several cancers, cardiovascular conditions, and age-related macular degeneration, confirming their healthy nutritional value. In scientific research, tomatoes have been employed as a model fruit, playing a pivotal role in studies on fruit development.

[0004] Assessing tomato quality is crucial for ensuring nutritious and flavorful produce, supporting the food industry, aiding farmers, and advancing scientific research on fruit quality development. The quality of the tomato fruit is primarily determined by a set of quality parameters, including firmness, total soluble solids (TSS), pH levels, unit size and weight, lycopene content, and citric and ascorbic acid levels. Traditional methods, such as the Folin reagent method and the Hart and Scott method, are used to estimate ascorbic acid and lycopene content, respectively, in the lab. However, these methods have several drawbacks. They require destruction of the tested fruit, which limits them to post-harvest assessment. They are also costly, labor-intensive, timeconsuming, and generate waste. Innovative techniques have been deployed to overcome such limitations, which include non-destructive proximal and remote sensing methods toassess tomato quality parameters. These sensing techniques include electronic smell sensors, spectroscopy, and the use of acoustic sensors.

[0005] Developing predictive models based on hyperspectral data (a branch of spectroscopy) is one such approach. Spectral-based indices, which are mathematical expressions of reflectance values derived at different spectral bands, may be used to assess various quality parameters such as the fruit’s color, its dry matter, firmness, TSS, pH, citric acid, lycopene content, and P-carotene contents. In this regard, the use of mobile or handheld hyperspectral equipment could be leveraged to assess pre-harvest fruit quality in the field. However, despite the potential of these approaches, their application in practice is limited due to the high cost of the commercial full-range hyperspectral cameras.

[0006] Accordingly, it would be beneficial to develop a machine learning-based tool that could be used to assess pre-harvest fruit quality, using affordable commercially-available cameras with a limited number of spectral bands.

[0007] The foregoing examples of the related art and limitations related therewith are intended to be illustrative and not exclusive. Other limitations of the related art will become apparent to those of skill in the art upon a reading of the specification and a study of the figures.SUMMARY OF INVENTION

[0008] The following embodiments and aspects thereof are described and illustrated in conjunction with systems, tools and methods which are meant to be exemplary and illustrative, not limiting in scope.

[0009] There is provided, in an embodiment, a computer-implemented method comprising: receiving, as input, spectral image data acquired with respect to a unit of tomato; calculating, from the spectral image data, spectral reflectance values associated with the unit of tomato in at least one specified pairwise combination of spectral bands; deriving, from the spectral reflectance values, a normalized difference spectral index (NDSI) value; and determining, based on the NDSI value, a value of at least one specified quality parameter associated with the unit of tomato.

[0010] There is also provided, in an embodiment, a system comprising at least one hardware processor; and a non-transitory computer-readable storage medium havingprogram instructions embodied therewith, the program instructions executable by the at least one hardware processor to: receive, as input, spectral image data acquired with respect to a unit of tomato, calculate, from the spectral image data, spectral reflectance values associated with the unit of tomato in at least one specified pairwise combination of spectral bands, derive, from the spectral reflectance values, a normalized difference spectral index (NDSI) value, and determine, based on the NDSI value, a value of at least one specified quality parameter associated with the unit of tomato.

[0011] There is further provided, in an embodiment, a computer program product comprising a non-transitory computer-readable storage medium having program instructions embodied therewith, the program instructions executable by at least one hardware processor to: receive, as input, spectral image data acquired with respect to a unit of tomato; calculate, from the spectral image data, spectral reflectance values associated with the unit of tomato in at least one specified pairwise combination of spectral bands; derive, from the spectral reflectance values, a normalized difference spectral index (NDSI) value; and determine, based on the NDSI value, a value of at least one specified quality parameter associated with the unit of tomato.

[0012] In some embodiments, the at least one specified combination of spectral bands is selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.

[0013] In some embodiments, the at least one specified combination of spectral bands is 733.0 nm and 902.4 nm, and the at least one specified quality parameter is selected from the group consisting of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

[0014] In some embodiments, the at least one specified combination of spectral bands is 649.9 nm and 733.0 nm, and the at least one specified quality parameter is selected from the group consisting of the following quality parameters: TSS and citric acid.

[0015] In some embodiments, the at least one specified combination of spectral bands is 649.9 nm and 902.4 nm, and the at least one specified quality parameter is TSS.

[0016] In some embodiments, the at least one specified combination of spectral bands is 706.3 nm and 519.1 nm, and the at least one specified quality parameter is pH.

[0017] In some embodiments, the at least one specified combination of spectral bands is 706.3 nm and 733.0 nm, and the at least one specified quality parameter is ascorbic acid.

[0018] In some embodiments, the NDSI value with respect to a specified pairwise combination of spectral bands (b1, b2is calculated based on the formula NDSI =

[0019] In some embodiments, the unit of tomato is a pre-harvest fruit.

[0020] In some embodiments, the spectral image data comprises data in no more than 10 spectral bands.

[0021] In some embodiments, the spectral image data comprises data in no more than 5 spectral bands.

[0022] In some embodiments, the spectral image data comprises data in no more than 2 spectral bands.

[0023] There is further provided, in an embodiment, a computer-implemented method comprising: receiving, as input, hyperspectral image data acquired with respect to a plurality of units of tomato; calculating spectral reflectance values with respect to each unit in each of the spectral bands represented in the hyperspectral image data; measuring ground-truth values of a set of quality parameters with respect to each of the units; constructing a training dataset comprising (i) features representing the spectral reflectance values with respect to each of the units, wherein the features are labeled with labels indicating the corresponding ground-truth values with respect to each of the unit, and (ii) the identities of combinations of spectral bands selected from a subset of the spectral bands represented in the hyperspectral image data, wherein each of the combinations is known as predictive with respect to at least one of the quality parameters; and training a machine learning model on the constructed training dataset, to obtain a trained machine learning model configured to be applied to spectral reflectance values measured with respect to a target unit of tomato in only the spectral bands included in the subset, to output a prediction with respect to the set of quality parameters in the target unit.

[0024] In some embodiments, the plurality of units of tomato comprises units representing two or more cultivars and different stages of maturation and ripeness.

[0025] In some embodiments, the target unit is a pre-harvest fruit.

[0026] In some embodiments, the set of quality parameters comprises: size, weight, firmness, total soluble solids (TSS), citric acid, ascorbic acid, lycopene content, and pH.

[0027] In some embodiments, the training dataset further comprises features representing a size of each of the units, and a Green -Red Vegetation Index value with respect to each of the units.

[0028] In some embodiments, the subset of spectral bands comprises the following spectral bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, and 828 nm.

[0029] In some embodiments, the combinations of spectral bands are selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.[00.30] In some embodiments, the combination of spectral band comprising 733.0 nm and 902.4 nm is predictive with respect to at least one of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

[0031] In some embodiments, the combination of spectral band comprising 649.9 nm and 733.0 nm is predictive with respect to at least one of the following quality parameters: TSS and citric acid.

[0032] In some embodiments, the combination of spectral band comprising 649.9 nm and 902.4 nm is predictive with respect to the following quality parameter: TSS.

[0033] In some embodiments, the combination of spectral band comprising 706.3 nm and 519.1 nm is predictive with respect to the following quality parameter: pH.

[0034] In some embodiments, the combination of spectral band comprising 706.3 nm and 733.0 nm is predictive with respect to the following quality parameter: ascorbic acid.

[0035] In addition to the exemplary aspects and embodiments described above, further aspects and embodiments will become apparent by reference to the figures and by study of the following detailed description.BRIEF DESCRIPTION OF THE FIGURES

[0036] FIG. l isa block diagram of an exemplary system which provides for prediction of pre-harvest fruit quality, based, at least in part, on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0037] FIG. 2A illustrates the functional steps in a method for predicting pre-harvest fruit quality, based on identifying commonly -used spectral indices that are predictive of quality parameters.

[0038] FIG. 2B is a schematic representation of the functional steps in a method for predicting pre-harvest fruit quality, based on identifying commonly-used spectral indices that are predictive of quality parameters.

[0039] FIG. 3A illustrates the functional steps in a method for predicting pre-harvest fruit quality based on NDSI modeling in a limited number of spectral bands.

[0040] Fig. 3B is a schematic representation of the functional steps in a method for predicting pre-harvest fruit quality based on NDSI modeling in a limited number of spectral bands.

[0041] FIG. 4A illustrates the functional steps in a method for training a machine learning model configured predict pre-harvest fruit quality, based on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0042] FIG. 4B is a schematic representation of the functional steps in a method for training a machine learning model configured predict pre-harvest fruit quality, based on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0043] FIG. 5 shows the best scoring NDSI combinations of spectral bands, and the quality parameter(s) predicted by each such NDSI combinations.

[0044] FIGS. 6A-6C show the spectral signature of the fruits in the input batch, from 400 nm to 1000 nm, of each fruit per cultivar, as well as for the average of all fruits per cultivar (black line).

[0045] FIG. 6D shows the average spectral signature of each cultivar plotted together.

[0046] FIGS. 7A-7B show the correlation coefficients (R2) and relative root mean square errors (rRMSE) of the linear regressions between five commonly-used spectral indices against the and quality parameters.

[0047] FIGS. 8A-8E show correlation maps between all NDSI combinations and TSS, citric acid, ascorbic acid, lycopene content, weight, firmness, and pH.DETAILED DESCRIPTION

[0048] Disclosed herein is a technique, embodied in a system, computer-implemented method, and computer program product, which provides for prediction of pre-harvest fruit quality, based, at least in part, on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0049] Spectral reflectance analysis is an important modality with a growing interest and a range of use cases. However, practical, field-based applications of hyperspectral imaging are limited by the fact that many existing hyperspectral imaging devices are physically large, complicated and expensive.

[0050] Accordingly, in some embodiments, the present technique is intended for use in conjunction with affordable multispectral cameras configured to capture a limited number of spectral bands. The present technique offers a low-cost practical solution, which could be adapted for use in mobile or handheld applications, to assess the quality of fruit in the field, at the pre-harvest stage.

[0051] The present technique provides immediate, non-contact, field-based prediction results, without the need to wait for lab results, or to subject the fruit to destructive analyses. Thus, the present technique can be adapted for use in the field by growers. In one example, the present technique may be implemented as part of automated (robotic) fruit pickers. The present technique can also be adapted for use by researchers, government and regulatory agents, the food industry, as well as at home or in stores by consumers.

[0052] The following discussion will focus on practical applications of the present technique with respect to predicting pre-harvest fruit quality in tomatoes. However, the principles of the present technique may be equally applicable in the case of other fruits and vegetables.

[0053] As used herein, “pre-harvest” refers to the period of growth of a fruit (such as tomato) before picking. The pre-harvest stage may encompass several maturation and ripening stages, including the mature green stage (color is visible inside but not outside), breaker stage (first blush, pink / red at blossom end), turning stage (10-30% red / orange), pink (30-60% colored), light red (60-90% colored), or fully ripe (90-100% red).

[0054] As used herein, the term “spectral band” refers to a contiguous range of wavelengths within the electromagnetic spectrum. Spectral bands can be characterizedby their center wavelength, bandwidth (full width at half maximum, FWHM), and spectral resolution (ability to distinguish adjacent bands).

[0055] As used herein, the terms “hyperspectral” or “hyperspectral imaging” refer generally to a remote sensing technique that captures and processes image data across a continuous range of wavelengths in the electromagnetic spectrum, typically producing hundreds of narrow, contiguous spectral bands (often 5-10 nm wide) per pixel. Hyperspectral imaging may cover the visible (VIS, 400-700 nm), near-infrared (NIR, 700-1,100 nm), shortwave-infrared (SWIR, 1,100-2,500 nm), mid-wave infrared (MWIR, 3,000-5,000 nm), and long-wave infrared (LWIR, 8,000-14,000 nm) spectral regions. Hyperspectral imaging generally provides high spectral resolution, typically smaller than 10 nm. Hyperspectral imaging thus generates a complete spectral signature for each pixel, forming a three-dimensional ‘data cube’ comprising two spatial dimensions plus the spectral dimension. This enables precise material identification and classification based on unique reflectance, absorbance, or emission properties.

[0056] As used herein, the terms “multispectral” or “multispectral imaging” generally refer to a remote sensing technique that captures image data at specific wavelength ranges across the electromagnetic spectrum, typically using fewer bands than hyperspectral imaging, e.g., 3 to 25 discrete spectral bands.

[0057] Plant spectral reflectance analysis techniques are based on optical reflectance and absorption properties of plant matter. In general, the metrics of spectral reflectance received from biological matter are dependent on the optical properties of the captured objects. Hence, the spectral reflectance received from plants may depend on optical properties of the plant, with particular regard to properties related to light absorption and scattering. For example, when a light beam having a specific intensity and wavelength is radiated at a biological object, part of this light beam is diffusely reflected from the surface of the object, while another part of the light beam passes through the surface into the tissue of the object and distributes there by means of multiple scattering. A fraction of this light scattered in the tissue exits back out from the surface as visible scattered light, whereby the intensity of this scattered light depends on the distance of the exit point from the irradiation point as well as on the wavelength of the light radiated. This dependence is caused by the optical properties of the biological matter. For example, different spectral bands (with different wavelengths) of the spectrum have different absorption levels in biological tissue. Thus, different absorption levels of differentwavelengths can lead to different metrics of spectral reflectance. These unique optical properties may be associated with physical (e.g., external and / or visual) and / or internal compositional attributes of the object, and therefore can be used for sensing, detection, and monitoring purposes.

[0058] In some embodiments, the present technique provides for predicting one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters of a pre-harvest tomato, including, but not limited to:- Physical external and / or visual criteria:o Color: Color in tomatoes is a primary indicator of ripeness and quality.Tomatoes should have a uniform orange-red to deep red color, depending on variety.o Firmness: Tomatoes should be firm but not hard. Tomato firmness affects handling, transportability, and shelflife.o Weight size and shape: Uniformity in size and a shape characteristic of a tomato variety are important for marketability and packaging.- Internal compositional criteria:o Total Soluble Solids (TSS): TSS is expressed as percentage Brix and measures the sugar content (mainly fructose and glucose), which is an indicator of sweetness.o Acidity (pH): Measures the organic acid content. Acidity in tomatoes provides tartness which balances the sweetness of the tomato.o Lycopene content: A powerful antioxidant found in tomatoes. o Citric acid levels: Another measure of acidity in tomatoes.o Ascorbic acid (vitamin C) levels: Vitamin C is essential for human health, and is also a powerful antioxidant.

[0059] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit, wherein the spectral reflectance is acquired in a limited number of spectral bands.

[0060] In some embodiments, the present technique provides for a prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit, wherein the spectral reflectance is acquired in no more than 10 specified spectral bands.

[0061] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit, wherein the spectral reflectance is acquired in no more than 5 specified spectral bands.

[0062] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit, wherein the spectral reflectance is acquired in 2 specified spectral bands.

[0063] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit, wherein the spectral reflectance is acquired in between 2 and 10 specified spectral bands, e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10 specified spectral bands.

[0064] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit in two or more of the following set of spectral bands: 902.4 nm, 733.0 nm, 706.3 nm, 649 nm, and 519.1 nm. In some embodiments, each of the spectral bands represents a band having a width of between 5-20 nm full width at half maximum (FWHM). In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in these spectral bands include weight, firmness, TSS, pH, lycopene content, citric acid, and / or ascorbic acid.

[0065] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional qualityparameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using a normalized difference spectral index (NDSI) modeling approach. The NDSI modeling comprises calculating spectral index values for specified combinations of spectral bands. For example, in the case of spectral bands> the NDSI is calculated as follows:

[0066] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using the following NDSI value in the following combination of spectral bands: 733.0 nm and 902.4 nm. In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in this spectral band combination include weight, firmness, TSS, and / or lycopene content.

[0067] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using the following NDSI value in the following combination of spectral bands: 649.9 nm and 902.4 nm. In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in this spectral band combination include TSS.

[0068] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using the following NDSI value in the following combination of spectral bands: 649.9 nm and 733.0 nm. In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in this spectral band combination include TSS and citric acid.

[0069] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional qualityparameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using the following NDSI value in the following combination of spectral bands: 519.1 nm and 706.3 nm. In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in this spectral band combination include pH.

[0070] In some embodiments, the present technique provides for prediction of one or more physical (e.g., external and / or visual) and / or internal compositional quality parameters in pre-harvest fruit, based, at least in part, on analyzing spectral reflectance properties of the fruit using the following NDSI value in the following combination of spectral bands: 706.3 nm and 733.0 nm. In some embodiments, the physical (e.g., external and / or visual) and / or internal compositional quality parameters predicted based on analyzing spectral reflectance properties of the fruit in this spectral band combination include ascorbic acid.

[0071] Reference is made to FIG. 1, which is a block diagram of an exemplary system 100 for realizing the present technique for prediction of pre-harvest fruit quality, based, at least in part, on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands. In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in between 2 and 10 specified spectral bands, e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10 specified spectral bands. In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in 2, 3, 4, or 5 spectral bands.

[0072] In some embodiments, system 100 may comprise a processing module 102, and a random-access memory (RAM) 104, and / or one or more non-transitory computer-readable storage device 106.

[0073] Processing module 102 may include components such as, but not limited to, one or more central processing units (CPUs), graphics processing units (GPUs), or any other suitable multi-purpose or specific processors or controllers. Processing module 102 may be operationally directly and / or indirectly connected to, and control the operation of, storage device 106 and all other components of system 100.

[0074] Storage device 106 may be or may include, for example, one or more non-transitory computer-readable storage device(s), a Random Access Memory (RAM), aread only memory (ROM), a Dynamic RAM (DRAM), a Synchronous DRAM (SDRAM), a double data rate (DDR) memory chip, a Flash memory, a volatile memory, a non-volatile memory, a cache memory, a buffer, a short term memory unit, a long term memory unit, or other suitable memory units or storage units.

[0075] In some embodiments, system 100 may store in storage device 106 software instructions or components configured to operate a processing unit (also ‘hardware processor,’ ‘CPU,’ or simply ‘processor’), such as processing module 102. The software instructions may be any executable code, e.g., a software application, a program, a process, task or script. In some embodiments, the software components may include an operating system, including various software components and / or drivers for controlling and managing general system tasks (e.g., memory management, storage device control, power management, etc.) and facilitating communication between various hardware and software components.

[0076] The software instructions and / or components operating processing module 102 may comprise one or more modules, such as an image data acquisition module 108, an image data processing module 110, a fruit quality analysis module 112, a spectral analysis module 114, a prediction model 116, and a user interface 118. These modules may be implemented in hardware only, software only, or a combination of both hardware and software.

[0077] In one example, system 100 may be realized as a portable or hand-held device configured to assess tomato quality in real time using multispectral imaging. The device may be used, for example, by field operators to assess tomatoes on vine, or by packing line operators to scan fruit. The device may be designed as a ruggedized, field-ready device, and include a user interface 118, such as a touchscreen-based interface. The device may include a multispectral imaging sensor configured to acquire spectral image data in 2-10 spectral bands. The device may include onboard edge Al capabilities for real-time low-latency inference and offline operation. The device may also have remote (e.g., cloud-based) Al capabilities for re-training, ensemble voting, or historical benchmarking purposes.

[0078] Image data acquisition module 108 may be configured to acquire and / or receive image data representing multiple imaging modalities, including RGB image data, multispectral image data, and hyperspectral image data.

[0079] Image data processing module 110 may employ any one or more suitable imaging analysis processes or techniques with respect to image data captured or received by image data acquisition module 108, or by any other interior and / or external imaging device.

[0080] Fruit quality analysis module 112 may be configured to obtain and / or receive analysis results of with respect to multiple physical and compositional properties of units of fruits.

[0081] Spectral analysis module 114 may be configured to process spectral data which may vary in aspects and properties, including with respect to the number of spectral bands and / or wavelengths represented therein, resolution, frame rate, format, and protocol.

[0082] Perdition module 116 may comprise any one or more suitable analysis and prediction algorithms or machine learning models. In some embodiments, prediction model 116 may be configured to predict pre-harvest fruit quality, based, at least in part, on spectral reflectance analysis results from spectral image data captured in a limited number of spectral bands. In some embodiments, prediction model 116 may be configured to be inferenced on an input spectral image data of a target pre-harvest fruit, captured in a limited number of spectral bands, to output a prediction with respect to the quality of the target fruit across a set of physical and compositional criteria.

[0083] System 100 as described herein is only an exemplary embodiment of the present invention, and in practice may be implemented in hardware only, software only, or a combination of both hardware and software. System 100 may have more or fewer components and modules than shown, may combine two or more of the components, or may have a different configuration or arrangement of the components. System 100 may include any additional component enabling it to function as an operable computer system, such as a motherboard, data busses, power supply, a network interface card, a display, an input device (e.g., keyboard, pointing device, touch-sensitive display), etc. (not shown).

[0084] The instructions of system 100 will now be discussed with reference to the flowchart of FIG. 2A, which illustrates the functional steps in a method 200 for predicting pre-harvest fruit quality, based, at least in part, on identifying commonly-used spectral indices that are predictive of quality parameters.

[0085] FIG. 2B is a schematic representation of a the various steps in method 200 for predicting pre-harvest fruit quality, based, at least in part, on identifying commonly -used spectral indices that are predictive of quality parameters.

[0086] The various steps of method 200 will be described with continuous reference to exemplary system 100 shown in FIG. 1. The various steps of method 200 may either be performed in the order they are presented or in a different order (or even in parallel), as long as the order allows for a necessary input to a certain step to be obtained from an output of an earlier step. In addition, the steps of method 200 may be performed automatically (e.g., by system 100 of FIG. 1), unless specifically stated otherwise. In addition, the steps of method 200 are set forth for exemplary purposes, and it is expected that modifications to the flowchart may be implemented as necessary or desirable.

[0087] Method 200 begins at step 202, system 100 executes image data acquisition module 108 to obtain or receive, as input, hyperspectral image data acquired by imaging an input batch of units of a specified fruit type. In some embodiments, the hyperspectral image data may represent one or more imaging instances of each of the units of the specified fruit type. In some embodiments, the hyperspectral image data may represent one or more imaging instances of each of the units of the specified fruit type, e.g., taken from different perspectives or sides of the fruit.

[0088] In some embodiments, system 100 may further execute image data acquisition module 108 to obtain or receive, as input, RGB or similar image data corresponding to the input hyperspectral image data.

[0089] In some embodiments, the imaged units of the specified type of fruit may comprise two or more cultivars of the specified fruit type, e.g., between 2-10 cultivars. In some embodiments, the imaged units of the specified fruit type may represent units of fruit in multiple different stages of ripeness. In some embodiments, the imaged units of the specified fruit type may represent units of fruit in all stages of the ripening process.

[0090] In some embodiments, the hyperspectral image data of the units of fruit may be obtained using a hyperspectral imaging system capable of capturing hyperspectral image data from a large number of narrow, contiguous bands across a wide spectral range.

[0091] In an exemplary implementation of step 202 by the present inventors, the received hyperspectral image data is associated with an experimental input batch comprising a total number of 567 fruits, in six sets of 40 to 60 fruits representing eachfive different cultivars. The fruits were collected from open fields, greenhouses, and net houses. The fruits in the input batch were selected to represent the full range of ripening stages (as described above), based on an assessment of external visual properties, such as some or all of the following stages:- Mature green stage (full size, firm, shiny green).- Breaker stage (first visible blush, pink / red at blossom end).Turning stage (10-30% red / orange).Pink (30-60% colored).- Light red (60-90% colored).Fully ripe (90-100% red).

[0092] The selected tomato cultivars included:- AS-200: This cultivar belongs to the determinate Roma segment and has a broad spectrum of disease resistance, performing optimally in outdoor conditions.- Eliot: An indeterminate single beef tomato cultivar with a typical flat shape and high disease resistance.- Shani: An indeterminate Cluster Tomato cultivar with a medium fruit size.- Round Robin: This cultivar belong to the determinate cherry dwarf tomatoes segment, and is characterized by green and red vertical stripes.- Mary's Cherry: This cultivar belong to the determinate cherry dwarf tomatoes segment, and is characterized by a slightly oblate shape and purple color.

[0093] Table 1 below summarizes main characteristics of the fruits collected for the experimental study, including set number, sample size (N), type, name of the cultivar, and conditions.Table 1: Experimental study composition.

[0094] In the exemplary implementation by the present inventors, the hyperspectral imaging of the units of tomato was obtained using SpecimlQ hyperspectral imaging system. The hyperspectral imaging system captures 204 wavelength bands on the spectrum ranging from 400 nm to 1000 nm, with a 7 nm FWHM bandwidth and a field-of-view (FOV) of 0.55 by 0.55 m at 1 m, and a spatial resolution of 512 x 512 pixels. The hyperspectral imaging system was calibrated for each successive image, using a 10x10 cm white reference calibration panel (90% reflectance) located near the imaged fruit to enable transforming radiance into relative reflectance values. To minimize the noise from background interference, the imaged fruit were placed on a black fabric cloth, wherein a full-spectrum projector was positioned as white reference (WR) in the scene. The WR was positioned at a height of 2 meters and directed downward towards the imaged fruit. Images were then taken with the hyperspectral imaging system fixed on a tripod at a diagonal distance of 1.5 m from the imaged fruit. An exposure time of 90-100 milliseconds was set to prevent over saturation in the image.

[0095] With reference back to FIG. 2A, in step 204, system 100 executes image data processing module 110 to perform one or more image processing steps with respect to the received hyperspectral image data, including:- Detecting individual units of fruit in the received image data.Calculating the size of the detected units of fruit in the image data.Segmenting and identifying an area-of-interest (AOI) associated with each of the detected units of fruit in the image data.- Removal of background areas in the image data.- Removal of white spots in the image data.

[0096] In some embodiments, at least some of the image processing steps may be performed on RGB image data received or acquired in step 202. In some embodiments, the RGB image data may be derived from the hyperspectral image data received in step 202.

[0097] Accordingly, in some embodiments, system 100 executes image data processing module 110 to detect one or more units of fruit in the image data obtained in step 202. In some embodiments, fruit detection may be performed using any suitable object detection method or technique.

[0098] In some embodiments, system 100 executes image data processing module 110 to segment each of the detected units of fruit in the image data, to identify and segment an AOI associated with the central area of the fruit. In some embodiments, the segmented AOI may be represented as a bounding box enclosing the detected fruit, another bounding region enclosing the detected fruit, and / or an outline of the detected fruit. In some embodiments, fruit AOI segmentation may be performed using any suitable object segmentation method or technique.

[0099] In some embodiments, system 100 executes image data processing module 110 to apply one or more further image processing steps, e.g., to remove the background in each of the image data, and / or to employ a threshold-based method to eliminate white spots on the imaged fruit. In some embodiments, white spot removal may be performed, e.g., by converting the received hyperspectral image data to RGB grayscale image data, and applying a thresholding method thereto.

[0100] In some embodiments, system 100 executes image data processing module 110 to calculate the size (i.e., AOI surface area) of each of the units of fruit represented in the received image data. In some embodiments, fruit size calculation may represent the calculated surface area of each of the units of fruit, and may be calculated using the known dimensions of the white reference (WR) in the scene (as detailed with reference to step 202). First, a pixel size may be calculated, based on the number of pixels that comprise the WR area in each image, as follows:. . WR pixel countpixel size = — - - - .WR known area (cm2)

[0101] Based on the calculated pixel size, the number of pixels included within the AOI representing each unit of fruit can be counted and converted into a total area value. In the case of units of fruit imaged from more the one side or aspect, the same calculation may be repeated with respect to each of the image data of the particular unit of fruit, and the average size of all the calculations may be used.

[0102] With reference back to FIG. 2A, in step 206, system 100 executes image data processing module 110 to measure spectral reflectance values (e.g., spectral mean and spectral standard deviation) for the AOIs associated with each of the units of fruit as detected in step 204. In some embodiments, spectral reflectance values are calculated with respect to each of the spectral bands captured in step 202. In the exemplary implementation by the present inventors, spectral reflectance for the identified AOIs associated with each of the units of fruit in the received image data, were calculated with respect to the 204 spectral bands captured by the hyperspectral imaging system, as described with reference to step 202 hereinabove.

[0103] In step 208, system 100 executes fruit quality analysis module 112 to obtain and / or determine ground-truth values associated with a plurality of physical (e.g., external and / or visual) and / or internal compositional quality parameters with respect to each of the plurality of units of fruit in the input batch described in step 202.

[0104] Accordingly, system 100 executes fruit quality analysis module 112 to obtain and / or determine ground-truth values associated with a plurality of physical quality parameters with respect to each of the units of fruit in the input batch, including, but not limited to, the following physical quality parameters:- Weight: The weight of each unit of fruit may be determined using a scale.- Firmness: The firmness of each unit of fruit may be measured using a durometer hardness, wherein measurements may be taken in one or more spots of each fruit, e.g., one or more sides near the equator. The measurement results may be then averaged for each unit of fruit.

[0105] In some embodiments, system 100 executes fruit quality analysis module 112 to obtain and / or determine ground-truth values associated with a plurality of compositional quality parameters with respect to each of the units of fruit in the input batch. The compositional quality assessment may be conducted by cutting the fruits along the longitudinal axis between the two poles, to ensure that each section contained a proportional representation of all tissue types and regions within the fruit, thus enabling multiple different analyses on each fruit. The compositional quality parameters include, but are not limited to, the following:- Total soluble solids (TSS): TSS may be expressed as percentage Brix (°Brix), and may be determined using a digital refractometer. The procedure may includefreezing the fruit, squeezing the defrosted fruit flesh through with a 3 -layer sterile gauze, to obtain the pipetted juice, pouring it onto the digital refractometer, and recording the measurement. Each measurement may be repeated multiple times, and the average value used as the TSS value for the unit of fruit.- Citric acid: A volume of 1 mL of juice from each unit of fruit may be mixed with 40 mL of double distilled water (DDW). Citric acid may then be determined using an electronic titration device.- Ascorbic acid: Ascorbic acid may be measured using the Folin reagent method.Fruit tissue (200 mg) may be grounded and mixed with 0.8 ml 10% (w / v) trichloroacetic acid and centrifuged for 20 min at 13,000 rpm under 4°C. The supernatant is then transferred to a new tube. Then, the reaction mixture is made with 160 pl and 20 pl of 10-fold diluted Folin reagent, and mixed with 40 pL of the fruit extract on a 96-well plate. The reaction mixture is incubated at room temperature in the dark for 10 min, and the absorbance is measured in a spectrophotometer at 760 nm. The ascorbate concentration per 100g is calculated to a standard calibration curve prepared from commercial L-ascorbic acid.pH: A volume of 1 ml of juice from each unit of fruit may be mixed with 40 mL of double distilled water (DDW). pH may then be determined using an electronic titration device.- Lycopene content: Lycopene measurements may be conducted by extracting fresh tissue samples according to Hart & Scott’s method. Fresh tissues of about 0.5 g is crushed with liquid nitrogen in a grinder. Carotenoids (including lycopene) are then extracted using 8 mL of Hexane: Acetone: Ethanol (50:25:25) solution in two steps of 4 mL each. The solution is mixed and vortexed. The resulting solution is mixed in a shaker for 5 min (vortex in the middle). One mL of saline solution (25% NaCl) is added, mixed, and vortexed, after which 8 mL of DDW is added and mixed. The solution is maintained undisturbed for about 10 min. When a visible separation of the phase is observed, the upper 2 mL of the phase (hexane and carotenoids) are collected and placed in a clean glassed tube. The hexane in the collected layer is then evaporated with nitrogen gas. The remaining carotenoids in the tube are mixed with an acetone-ethanol (50:50) solution of 1 mL. Samples are then mixed and vortexed. From the final solutionof carotenoids in the acetone-ethanol solution, 70 pL is placed on a 96-well plate that is resistant to organic compounds and is compatible with UV measurements, having three repetitions per sample. The absorbance of the extract is measured at 503 nm to determine the lycopene concentration as follows:1lycopene contentwhere A is the absorbance at 503 nm, V is the total extract volume in milliliters, p is the sample weight in grams, and A j °CI11is the lycopene extinction coefficient in petroleum ether, equal to 3450 (100 mL g-1cm-1).

[0106] With reference back to FIG. 2A, in step 210, system 100 executes spectral analysis module 114 to calculate linear regressions of commonly used spectral -based indices with respect to each unit of fruit in the input batch received in step 202. In some embodiments, the regressions are calculated based on the spectral values measured in step 206 against the ground-truth values of each of the physical (e.g., external and / or visual) and / or internal compositional quality parameter values determined in step 208.

[0107] Spectral indices are mathematical expressions of the reflectance values of two or more wavelength bands. The calculation of these indices is not sensitive to fluctuations in signal strength, given that most of them use ratio-based or normalized formulas, prioritizing the relationship between different wavelengths.

[0108] Accordingly, system 100 executes spectral analysis module 114 to perform linear regression calculations based on the spectral values measured in step 206 and the ground-truth values of the physical (e.g., external and / or visual) and / or internal compositional quality parameter values determined in step 208 with respect to the input batch received in step 202, using the formulations of five commonly-used spectral indices listed in Table 2 below.

[0109] In some embodiments, the spectral indices may be calculated from the spectral reflectance values calculated in step 206, to determine potential correlations with tomato quality parameters. The average spectral index value may be determined for each of the five indices listed in Table 2 below, with respect to each calculated AOI of each unit of fruit. The calculated index may then be linearly regressed against the ground-truth values of the quality parameters determined in step 208 with respect to the input batch received in step 202, wherein a regression score may be assigned to each calculation based on thecorrelation coefficient (R2) and / or root mean square error (RMSE) of the regression result.

[0110] Table 2 below provides the formulation of the five indices alongside their main characteristics and references.Table 2: The formulation and main characteristics of the five spectral indices used in this study to predict fruit quality parameters.&* G: Green (525 nm - 532 nm); R: Red (452 nm - 460 nm); B: Blue (627 nm - 635 nm).

[0111] In step 212, system 100 may configure prediction model 116 to use the indices identified as having the highest regression scores with respect to each of the quality parameters, as the indices most predictive of these one or more quality parameters.

[0112] The instructions of system 100 will now be discussed with reference to the flowchart of FIG. 3 A, which illustrates the functional steps in a method 220 for predicting pre-harvest fruit quality based on NDSI modeling in a limited number of spectral bands.

[0113] In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in between 2 and 10 specified spectral bands, e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10 specified spectral bands. In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in 2, 3, 4, or 5 spectral bands.

[0114] Fig. 3B is a schematic representation of the functional steps in a method for predicting pre-harvest fruit quality based on NDSI modeling in a limited number of spectral bands.

[0115] The various steps of method 220 will be described with continuous reference to exemplary system 100 shown in FIG. 1 and the method steps of FIGS. 2A and 3 A. The various steps of method 220 may either be performed in the order they are presented or in a different order (or even in parallel), as long as the order allows for a necessary input to a certain step to be obtained from an output of an earlier step. In addition, the steps of method 220 may be performed automatically (e.g., by system 100 of FIG. 1), unless specifically stated otherwise. In addition, the steps of method 220 are set forth for exemplary purposes, and it is expected that modifications to the flowchart may be implemented as necessary or desirable.

[0116] In step 222, system 100 performs steps 202-208 of method 200 as detailed hereinabove, which are incorporated herein by reference.

[0117] In step 224, system 100 executes spectral analysis module 114 to calculate normalized difference spectral index (NDSI) values for combinations of spectral bands represented in the hyperspectral image data received in step 202, based on the spectral reflectance values measured in step 206 with respect to each AOI of each unit of fruit in the input batch. For example, system 100 may execute spectral analysis module 114 to calculate normalized difference spectral index (NDSI) values for all possible pairwise combinations of spectral bands represented in the hyperspectral image data received in step 202.

[0118] In some embodiments, NDSI calculations use combinations of combinations of spectral bands as follows:where b and b2are the reflectance values of any two wavelength bands.

[0119] In the exemplary implementation by the present inventors described in step 202 of method 200, each hyperspectral image contains 204 bands in the spectral range from 400 to lOOOnm, for a total of 20,706 possible pairwise NDSI combinations.

[0120] In step 226, system 100 executes spectral analysis module 114 to calculate linear regressions between each NDSI value representing a combination of spectral bands, and the ground-truth values of each of the quality parameters determined in step 208 with respect to the input batch received in step 202, to generate a set of linear regression scores represented as the R2and / or the RMSE of the correlation calculations. Based on the regression scores, it is possible to identify those spectral band combinations which have the best predictive performance with respect to each of the quality parameters.

[0121] In one example, the regression scores are represented as R2values. The scores may be normalized by dividing them by the highest R2achieved for the best NDSI value correlation with the quality parameter. A normalized score map may be generated for each quality parameter with values from 0 to 1.

[0122] In another example, both R2and RMSE values may be combined into a single regression score, by generating a normalized combined score. The normalized combined score may be generated by normalizing the R2and RMSE values by their minimum and maximum values over all results, and then averaging both normalized scores to derive a single best-performing normalized score. Accordingly, the normalized scores (i.e., combined R2and RMSE values) were derived as follows:Normalized Scorewhere Scoremax is the maximum score value in the NDSI correlation map domain - i.e., the highest R2value or the lowest RMSE value, and Scoremin is the minimum score, which is the lowest R2or highest RMSE value. A normalized score map may be generated with score ranges from 1 for the best R2and RMSE values (highest and lowest, respectively), to 0 for the worst R2and RMSE values (lowest and highest, respectively).

[0123] In some embodiments, a Gaussian blur may be applied to the calculated regression scores, to facilitate data smoothing by reducing noise, improving thresholding for hotspot detection, and enhancing morphological analysis.

[0124] In step 228, system 100 executes spectral analysis module 114 to generate correlation ‘heatmaps’ with respect to the ground-truth values of each quality parameter, based on the calculated regression scores for all NDSI combinations. The heatmaps provide a visualization of the data using color to represent the magnitude of values. Each heatmap represents, for a single quality parameter, a 2-dimentioal space wherein the regression scores calculated in step 226 with respect to all pairwise NDSI combinations are represented by color to indicate magnitude of values.

[0125] In some embodiments, system 100 further executes spectral analysis module 114 to mark the heatmaps to indicate zones with the highest values (termed herein ‘hotspots’). These hotspots show the best NDSI combinations producing the highest correlations for each quality parameter. The hotspots are identified within each heatmap using percentile-based thresholding. The threshold may be adjusted to optimize the size of the hotspot zones, so as to exclude less informative NDSI combinations. In some embodiments, the identified hotspots may undergo morphological adjustments to fill minor gaps and prevent fragmented zones.

[0126] In step 230, system 100 executes spectral analysis module 114 to identify overlapping ‘hotspot1between the individual heatmaps generated in step 226 for each of the quality parameters. These overlap zones may identify NDSI combinations which are predictive or informative with respect to two or more quality parameters. This may help to minimizing the universe of combinations of spectral bands necessary to predict as many quality parameters as possible.

[0127] Accordingly, in some embodiments, system 100 executes spectral analysis module 114 to identify overlapping zones between two or more heatmaps (each representing a single quality parameter). These overlapping zones indicate spectral band combinations (i.e., NDSI combinations) which represent the most predictive or informative NDSI combinations with respect to two or more quality parameters, based on their regression correlation scores. In one example, system 100 executes spectral analysis module 114 to overlay the hotspots in the various heatmaps, to create an evaluation heatmap based on the average of the regression scores over all the heatmaps (wherein each heatmap represents a single quality parameter). This evaluation heatmap represents the average regression scores corresponding to all quality parameters having overlapping hotspot zones. For instance, the hotspots identified in the TSS and pH regression score heatmaps may be overlayed and the individual regression scoresaveraged. Then a new combined hotspot zone with the maximum averaged values may be derived, which represent the best performing NDSI combinations for both of TSS and pH. The maximum regression score within the combined hotspot zone in the overlayed evaluation heatmap is selected as the best NDSI combination for predicting the overlapping quality parameters.

[0128] In step 232, system 100 may configure prediction model 116 to use the best scoring NDSI combinations of spectral bands with respect to one or more quality parameters, as the NDSI combinations most predictive or informative of these one or more quality parameters.

[0129] Table 3 below shows the best scoring pairwise NDSI combinations of spectral bands identified in the experimental implementation by the present inventors, and the quality parameter(s) predicted by each such NDSI combination:Table 3:

[0130] FIG. 5 shows the best scoring NDSI combinations of spectral bands, and the quality parameter(s) predicted by each such NDSI combination.

[0131] The instructions of system 100 will now be discussed with reference to the flowchart of FIG. 4 A, which illustrates the functional steps in a method 240 for training a prediction model configured predict pre-harvest fruit quality, based on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0132] Fig. 4B is a schematic representation of the functional steps in a method for training a prediction model configured predict pre-harvest fruit quality, based on analyzing spectral reflectance properties of the fruit in a limited number of spectral bands.

[0133] In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in between 2 and 10 specified spectral bands, e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10 specified spectral bands. In some embodiments, the present technique is based, at least in part, on analyzing spectral reflectance properties of the fruit in 2, 3, 4, or 5 spectral bands.

[0134] The various steps of method 240 will be described with continuous reference to exemplary system 100 shown in FIG. 1 and the method steps of FIGS. 2A, 3 A and 4A. The various steps of method 240 may either be performed in the order they are presented or in a different order (or even in parallel), as long as the order allows for a necessary input to a certain step to be obtained from an output of an earlier step. In addition, the steps of method 240 may be performed automatically (e.g., by system 100 of FIG. 1), unless specifically stated otherwise. In addition, the steps of method 240 are set forth for exemplary purposes, and it is expected that modifications to the flowchart may be implemented as necessary or desirable.

[0135] In step 242, system 100 performs steps 202-208 of method 200 as detailed hereinabove, which are incorporated herein by reference.

[0136] In step 244, system 100 executes spectral analysis module 114 to construct a training dataset from features extracted from (i) the spectral reflectance values calculated in step 206, (ii) the ground-truth values of the physical (e.g., external and / or visual) and / or internal compositional quality parameter determined in step 208, and (iii) the NDSI combinations, calculated using method 220.

[0137] In some embodiments, the training dataset is designed to enable a machine learning model to predict tomato quality parameters using spectral reflectance features from a limited set of no more than 5 bands. The dataset emphasizes feature sparsity by incorporating explicit knowledge about predictive spectral band combinations with respect to each quality parameter, thus guiding the model to prioritize these combinations during training. This explicitly encodes domain knowledge (from the experimental results associated with best performing NDSI combinations as described with reference to method 220 hereinabove) to train the model to focus on informative pairs and minimize the total number of required bands while maintaining predictive accuracy.

[0138] In some embodiments, the training dataset comprises a plurality of sets of features extracted from the spectral reflectance values calculated in step 206 of method200, with respect to each unit of fruit in the input batch received in step 202. In some embodiments, the sets of features comprise one or more of the following feature categories:The best scoring NDSI combinations, determined as detailed hereinabove with reference to step 230 of method 220 and as reported in Table 3 above.Spectral reflectance values (e.g., spectral mean and spectral standard deviation) for each fruit unit in each of the spectral bands comprising the best NDSI combinations, determined as detailed hereinabove with reference to steps 224- 230 of method 220 and as reported in Table 3 hereinabove.AOI for each fruit unit, as determined in step 204 of method 200 hereinabove. GRVI (Green-Red Vegetation Index) value for each fruit unit, which provides an indication with respect to the ripeness stage of the fruit unit, as may be derived from RGB image data.

[0139] In some embodiments, the sets of features may be annotated with labels indicating ground-truth values of the quality parameters associated with the respective unit of fruit, e.g., values associated with one or more of size, weight, firmness, total soluble solids (TSS), citric acid, ascorbic acid, lycopene content, and / or pH.

[0140] In some embodiments the training dataset may be divided into training and test sets, with 70% of the data samples allocated to the training set, and the remaining 30% to the test set. In some embodiments, each of the training and test sets are balanced across tomato cultivars, ripeness stages, and growing conditions represented in the input batch.

[0141] In step 246, system 100 trains a machine learning model using the training dataset constructed in step 244.

[0142] In some embodiments, the machine learning model may be implemented as one or more of:- Artificial Neural Network (ANN) model: An ANN-based model is well suited for capturing non-linear, complex interactions.- Random Forest model: Random Forest is robust to noise and outliers.- Extreme gradient boosting (XGBoost) model: XGBoost handles sparse data and mixed feature types well.

[0143] In some embodiments, the machine learning model is an ensemble model integrating two or more distinct algorithms (e.g., ANN, Random Forest and XGBoost). Each algorithm processes the same input features, but leverages different learning paradigms, leading to varying performance across quality parameters. Thus, the ensemble may select the best-performing algorithm per quality parameter. This approach leverages individual model strengths and mitigates individual model weaknesses.

[0144] In step 248, system 100 may inference or apply the trained machine learning model, which may be realized in prediction model 116, on target spectral data acquired with respect to a target pre-harvest fruit.

[0145] The target spectral data comprises spectral reflectance values captured in a limited number of spectral bands, e.g., no more than 5 spectral bands. In some embodiments, the spectral bands represented in the target data are 519.1 nm, 649.9 nm, 706.3 nm, 733.0 nm, and 902.4 nm.

[0146] The trained machine learning model, which may be realized in prediction model 116, is then applied to the target data, to obtain predictions with respect to one or more of size, weight, firmness, total soluble solids (TSS), citric acid, ascorbic acid, lycopene content, and / or pH of the target fruit.Experimental Results

[0147] As noted, an exemplary implementation of methods 200, 220, and 240 disclosed hereinabove was conducted by the present inventors, using an experimental input batch comprising a total number of 567 fruits, in six sets of 40 to 60 fruits representing each five different cultivars. This characteristics of the input batch are described in detail with reference to step 202 of method 200.

[0148] FIGS. 6A-6C show the spectral signature ranges of the fruits in the input batch, from 400 nm to 1000 nm, of each fruit unit per cultivar, as well as for the average of all fruits per cultivar (thick black line). FIG. 6D shows the average spectral signature of all cultivars plotted together.Commonly-Used Spectral Indices Prediction

[0149] Of the commonly used spectral -based indices, it was found that firmness and lycopene were significantly correlated (R2>0.7; p<0.05) with GLI, LYC, INA, and GRVI,but not with ZMI. FIGS. 7A-7B show the correlation coefficients (R2) and relative root mean square errors (rRMSE) of the linear regressions between five commonly -used spectral indices (GLI, LYC, ZMI, INA, and GRVI - see formulations of these indices in Table 2 above), against the quality parameters (weight, firmness, TSS, citric and ascorbic acids, lycopene content, and pH).NDSI-Based Prediction

[0150] The NDSI approach bore similar results to the commonly used spectral indices, with the highest correlations for firmness and lycopene content. However, there were also NDSI combinations correlated with TSS, citric acid, ascorbic acid, weight, firmness, and pH. The maximum number of overlapping parameters was 4, yielding a combination of bands at ~650nm with a range of 700 - 990nm and another combination of ~710nm with a similar 700 - 950nm wavelength range. This allows predicting the seven quality parameters with as a few as 5 bands (with overlapping bands) and up to 10 bands (without overlapping, because two parameters had the best correlation with the same two-bands NDSI combination).

[0151] FIGS. 8A-8E show correlation maps between all NDSI combinations and the quality parameters (TSS, citric acid, ascorbic acid, lycopene content, weight, firmness, and pH). Hotspots in the maps are delimited with a white line, showing the best NDSI two-band combinations that produced the highest correlations with the quality parameter. The hotspot plot in FIG. 8E shows the number of overlapping hotspot maps, meaning that, for example, for 4 overlapped hotspots, the same NDSI combination could be used to derive four different quality parameters.Machine Learning Models

[0152] The present inventors experimented with a machine learning model based on the following algorithms and architectures:- ANN: An ANN model comprising two hidden layers: the first with 205 neurons and the second with 102 neurons. The ANN architecture applies the ReLU activation function for its ability to introduce non-linearity efficiently. The output layer employs a linear activation function which facilitates the regression nature of the present task. Hyperparameters for the ANN included a batch size of 32 and 200 epochs and selections made to optimize the training process. The Adamoptimizer was used for training, with the learning rate set at a default value of 0.01.- Random Forest: The Random Forest -based model has a number of estimators set to 100, to seek an optimal balance between computational efficiency and model performance. The maximum depth (max depth) was adjusted to 6, to prevent individual trees from growing excessively complex and overfitting to the training data. The random state was fixed at 42 to ensure reproducibility.- XGBoost: The XGBoost model also employed 100 estimators, aiming for an optimal balance between performance and computational demand. The maximum depth (max depth) was also limited to 6 to mitigate overfitting risks. A random state of 42 ensures consistent model behavior across executions.- MVR: the MVR approach, which operates without the need for hyperparameter tuning, served as a benchmark for comparison, leveraging its simplicity and interpretability to assess the effectiveness of the other models.

[0153] The following are the correlation coefficients ( / ?2) achieved by the machine learning models with respect to each quality parameter, using 5 spectral bands (519.1 nm, 649 nm, 706.3 nm, 733.0 nm, and 902.4 nm):- Weight: 0.95 (ANN).Firmness: 0.88 (RF).- TSS: 0.72 (RF).- Citric acid: 0.62 (RF).- Ascorbic acid: 0.35 (RF).- Lycopene content: 0.83 (ANN).- pH: 0.64 (RF).

[0154] The inventors further compared these results with predictions based on a greater number of spectral bands:6 bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, and 828 nm.7 bands: 519.1 nm, 649 nm, 706.3 nm, 902.4 nm, 828 nm, 872 nm, and 938 nm.8 bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, 828 nm, 872 nm, and 938 nm.9 bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, 620 nm, 828 nm, 872 nm, and 938 nm.10 bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, 471 nm, 620 nm, 828 nm, 872 nm, and 938 nm.

[0155] Using only 5 spectral bands (i.e., 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, and 828 nm) provided the best prediction results with respect to weight, TSS and citric acid. Using 7 bands provided only slightly better prediction results than 5 bands with respect to lycopene content (0.85 vs. 0.83) and pH (0.67 vs. 0.64). Using 9 bands provided only slightly better prediction results than 5 bands with respect to firmness content (0.89 vs. 0.88). Using 10 bands provided slightly better prediction results than 5 bands with respect to ascorbic acid (0.45 vs. 0.35).

[0156] The present invention may be a system, a method, and / or a computer program product. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.

[0157] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passingthrough a fiber-optic cable), or electrical signals transmitted through a wire. Rather, the computer readable storage medium is a non-transient (i.e., not-volatile) medium.

[0158] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0159] Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object-oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user’s computer, partly on the user’s computer, as a stand-alone software package, partly on the user’s computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user’s computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, a field-programmable gate array (FPGA), or a programmable logic array (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention. In some embodiments, electronic circuitry including, for example, an application-specific integrated circuit (ASIC), maybe incorporate the computer readable program instructions already at time of fabrication, such that the ASIC is configured to execute these instructions without programming.

[0160] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0161] These computer readable program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0162] The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0163] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructionsfor implementing the specified logical function(s). It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.

[0164] In the description and claims, each of the terms “substantially,” “essentially,” and forms thereof, when describing a numerical value, means up to a 20% deviation (namely, ±20%) from that value. Similarly, when such a term describes a numerical range, it means up to a 20% broader range - 10% over that explicit range and 10% below it).

[0165] In the description, any given numerical range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range, such that each such subrange and individual numerical value constitutes an embodiment of the invention. This applies regardless of the breadth of the range. For example, description of a range of integers from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., as well as individual numbers within that range, for example, 1, 4, and 6. Similarly, description of a range of fractions, for example from 0.6 to 1.1, should be considered to have specifically disclosed subranges such as from 0.6 to 0.9, from 0.7 to 1.1, from 0.9 to 1, from 0.8 to 0.9, from 0.6 to 1.1, from 1 to 1.1 etc., as well as individual numbers within that range, for example 0.7, 1, and 1.1.

[0166] The descriptions of the various embodiments of the present invention have been presented for purposes of illustration, but are not intended to be exhaustive or limited to the explicit descriptions. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

[0167] In the description and claims of the application, each of the words “comprise,” “include,” and “have,” as well as forms thereof, are not necessarily limited to members in a list with which the words may be associated.

[0168] Where there are inconsistencies between the description and any document incorporated by reference or otherwise relied upon, it is intended that the present description controls.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method comprising:receiving, as input, spectral image data acquired with respect to a unit of tomato; calculating, from said spectral image data, spectral reflectance values associated with said unit of tomato in at least one specified pairwise combination of spectral bands;deriving, from said spectral reflectance values, a normalized difference spectral index (ND SI) value; anddetermining, based on said NDSI value, a value of at least one specified quality parameter associated with said unit of tomato.

2. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.

3. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is 733.0 nm and 902.4 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

4. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is 649.9 nm and 733.0 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: TSS and citric acid.

5. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is 649.9 nm and 902.4 nm, and said at least one specified quality parameter is TSS.

6. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is 706.3 nm and 519.1 nm, and said at least one specified quality parameter is pH.

7. The computer-implemented method of claim 1, wherein said at least one specified combination of spectral bands is 706.3 nm and 733.0 nm, and said at least one specified quality parameter is ascorbic acid.

8. The computer-implemented method of any one of claims 1 -7, wherein said NDSI value with respect to a specified pairwise combination of spectral bands (&-£, b2) iscalculated based on the formula NDSI =bj+b29. The computer-implemented method of any one of claims 1-8, wherein said unit of tomato is a pre-harvest fruit.

10. The computer-implemented method of any one of claims 1-9, wherein said spectral image data comprises data in no more than 10 spectral bands.

11. The computer-implemented method of any one of claims 1-9, wherein said spectral image data comprises data in no more than 5 spectral bands.

12. The computer-implemented method of any one of claims 1-9, wherein said spectral image data comprises data in no more than 2 spectral bands.

13. A system comprising:at least one hardware processor; anda non-transitory computer-readable storage medium having program instructions embodied therewith, the program instructions executable by said at least one hardware processor to:receive, as input, spectral image data acquired with respect to a unit of tomato,calculate, from said spectral image data, spectral reflectance values associated with said unit of tomato in at least one specified pairwise combination of spectral bands,derive, from said spectral reflectance values, a normalized difference spectral index (NDSI) value, anddetermine, based on said NDSI value, a value of at least one specified quality parameter associated with said unit of tomato.

14. The system of claim 13, wherein said at least one specified combination of spectral bands is selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.

15. The system of claim 13, wherein said at least one specified combination of spectral bands is 733.0 nm and 902.4 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

16. The system of claim 13, wherein said at least one specified combination of spectral bands is 649.9 nm and 733.0 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: TSS and citric acid.

17. The system of claim 13, wherein said at least one specified combination of spectral bands is 649.9 nm and 902.4 nm, and said at least one specified quality parameter is TSS.

18. The system of claim 13, wherein said at least one specified combination of spectral bands is 706.3 nm and 519.1 nm, and said at least one specified quality parameter is pH.

19. The system of claim 13, wherein said at least one specified combination of spectral bands is 706.3 nm and 733.0 nm, and said at least one specified quality parameter is ascorbic acid.

20. The system of any one of claims 13-19, wherein said NDSI value with respect to a specified combination of spectral bands (bltb2) is calculated based on the formula21. The system of any one of claims 13-20, wherein said unit of tomato is a preharvest fruit.

22. The system of any one of claims 13-21, wherein said spectral image data comprises data in no more than 10 spectral bands.

23. The system of any one of claims 13-22, wherein said spectral image data comprises data in no more than 5 spectral bands.

24. The system of any one of claims 13-23, wherein said spectral image data comprises data in no more than 2 spectral bands.

25. A computer program product comprising a non-transitory computer-readable storage medium having program instructions embodied therewith, the program instructions executable by at least one hardware processor to:receive, as input, spectral image data acquired with respect to a unit of tomato; calculate, from said spectral image data, spectral reflectance values associated with said unit of tomato in at least one specified pairwise combination of spectral bands;derive, from said spectral reflectance values, a normalized difference spectral index (NDSI) value; anddetermine, based on said NDSI value, a value of at least one specified quality parameter associated with said unit of tomato.

26. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.

27. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is 733.0 nm and 902.4 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

28. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is 649.9 nm and 733.0 nm, and said at least one specified quality parameter is selected from the group consisting of the following quality parameters: TSS and citric acid.

29. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is 649.9 nm and 902.4 nm, and said at least one specified quality parameter is TSS.

30. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is 706.3 nm and 519.1 nm, and said at least one specified quality parameter is pH.

31. The computer program product of claim 25, wherein said at least one specified combination of spectral bands is 706.3 nm and 733.0 nm, and said at least one specified quality parameter is ascorbic acid.

32. The computer program product of any one of claims 25-31, wherein said NDSI value with respect to a specified combination of spectral bands (bltb2) is calculated based on the formula33. The computer program product of any one of claims 25-32, wherein said unit of tomato is a pre-harvest fruit.

34. The computer program product of any one of claims 25-33, wherein said spectral image data comprises data in no more than 10 spectral bands.

35. The computer program product of any one of claims 25-34, wherein said spectral image data comprises data in no more than 5 spectral bands.

36. The computer program product of any one of claims 25-35, wherein said spectral image data comprises data in no more than 2 spectral bands.

37. A computer-impl emented method compri sing :receiving, as input, hyperspectral image data acquired with respect to a plurality of units of tomato;calculating spectral reflectance values with respect to each unit in each of the spectral bands represented in the hyperspectral image data;measuring ground-truth values of a set of quality parameters with respect to each of the units;constructing a training dataset comprising (i) features representing said spectral reflectance values with respect to each of the units, wherein said features are labeled with labels indicating said corresponding ground-truth values with respect to each of the unit, and (ii) the identities of combinations of spectral bands selected from a subset of said spectral bands represented in the hyperspectral image data, whereineach of said combinations is known as predictive with respect to at least one of said quality parameters; andtraining a machine learning model on the constructed training dataset, to obtain a trained machine learning model configured to be applied to spectral reflectance values measured with respect to a target unit of tomato in only said spectral bands included in said subset, to output a prediction with respect to said set of quality parameters in said target unit.

38. The computer-implemented method of claim 37, wherein said plurality of units of tomato comprises units representing two or more cultivars and different stages of maturation and ripeness.

39. The computer-implemented method of any one of claims 37 or 38, wherein said target unit is a pre-harvest fruit.

40. The computer-implemented method of any one of claims 37-39, wherein said set of quality parameters comprises: size, weight, firmness, total soluble solids (TSS), citric acid, ascorbic acid, lycopene content, and pH.

41. The computer-implemented method of any one of claims 37-40, wherein said training dataset further comprises features representing a size of each of said units, and a Green-Red Vegetation Index value with respect to each of said units.

42. The computer-implemented method of any one of claims 37-41, wherein said subset of spectral bands comprises the following spectral bands: 519.1 nm, 649 nm, 706.3 nm, 733.0 nm, 902.4 nm, and 828 nm.

43. The computer-implemented method of any one of claims 37-42, wherein said combinations of spectral bands are selected from the group consisting of the following combinations of spectral bands: 733.0 nm and 902.4 nm; 649.9 and 733.0 nm; 649.9 nm and 902.4 nm; 706.3 nm and 519.1 nm; and 706.3 nm and 733.0.

44. The computer-implemented method of 43, wherein said combination of spectral band comprising 733.0 nm and 902.4 nm is predictive with respect to at least one of the following quality parameters: firmness, weight, total soluble solids (TSS), and lycopene content.

45. The computer-implemented method of 43, wherein said combination of spectral band comprising 649.9 nm and 733.0 nm is predictive with respect to at least one of the following quality parameters: TSS and citric acid.

46. The computer-implemented method of 43, wherein said combination of spectral band comprising 649.9 nm and 902.4 nm is predictive with respect to the following quality parameter: TSS.

47. The computer-implemented method of 43, wherein said combination of spectral band comprising 706.3 nm and 519.1 nm is predictive with respect to the following quality parameter: pH.

48. The computer-implemented method of 43, wherein said combination of spectral band comprising 706.3 nm and 733.0 nm is predictive with respect to the following quality parameter: ascorbic acid.

Images