Skip to main content

AUTOMATIZACION DE LA INSPECCION DE LA CALIDAD INTERNA Y SEGURIDAD DE FRUTAS EN LINEA, UTILIZANDO ESPECTROSCOPIA VIS/NIR E INTELIGENCIA ARTIFICIAL

PID2023-150192OR-C31

•
Funding agency name Agencia Estatal de Investigación
Funding agency acronym AEI
Program Programa Estatal para Impulsar la Investigación Científico-Técnica y su Transferencia
Subprogram Subprograma Estatal de Generación de Conocimiento
Call Proyectos de I+D+i (Generación de Conocimiento y Retos Investigación)
Call year 2023
State Plan 2021-2023 Plan Estatal de Investigación Científica y Técnica y de Innovación
Beneficiary centre INSTITUTO VALENCIANO DE INVESTIGACIONES AGRARIAS CAV

Resultados relacionados

Resultados totales (Incluyendo duplicados): 7
Encontrada(s) 1 página(s)

RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Tesis doctoral (DoctoralThesis)

Advanced Non-Destructive Evaluation of Fruit Quality and Safety Using Spectral Information and Artificial Neural Networks

  • Castillo Gironés, Salvador
[ES] Garantizar una buena calidad interna e inocuidad en los productos frutícolas es un gran desafío para la industria agroalimentaria, ya que los métodos tradicionales de evaluación suelen ser destructivos e ineficientes. Esta tesis doctoral tuvo como objetivo explorar el uso de dos técnicas ópticas no destructivas: imagen hiperespectral (HSI) y espectroscopía Raman, combinadas con modelos avanzados de redes neuronales, para predecir atributos de calidad interna, aceptación del consumidor y seguridad alimentaria en fresas, caquis, ciruelas y manzanas. La investigación incluyó cuatro estudios principales. Primero, se aplicaron HSI y redes neuronales a 17 cultivares de fresa para predecir parámetros clave de calidad como contenido de azúcar, acidez y textura, así como la aceptación general del consumidor e identificación de cultivares para prevenir fraudes alimentarios. Segundo, se usaron modelos de aprendizaje profundo basados en autoencoders para clasificar la imagen hiperespectral de caquis 'Rojo Brillante' en grupos de textura relevantes para la clasificación en centrales hortofrutícolas. Tercero, se entrenaron y compararon tres redes neuronales convolucionales (CNN) para detectar golpes invisibles en ciruelas 'Presenta', con el objetivo de reducir pérdidas poscosecha. Cuarto, la espectroscopía Raman combinada con CNN permitió la detección e identificación de múltiples compuestos activos de productos fitosanitarios en mezclas complejas comúnmente usadas en huertos de manzana. Los resultados demostraron que HSI y redes neuronales proporcionaron buenas predicciones para la calidad de la fresa (TSS R² = 0.85, TA R² = 0.81), aceptación media del consumidor por cultivar (R² = 0.78) y discriminación de cultivares (F1 = 84 %). En caquis, los autoencoders lograron una reducción del 99 % en la dimensionalidad de los datos y un 81.3 % de precisión en la clasificación de textura. Las CNN detectaron golpes en etapa temprana en ciruelas con una puntuación F1 del 90 % y la identificación de longitudes de onda clave. La espectroscopía Raman con CNNs identificó la presencia de fitosanitarios en mezclas complejas con hasta un F1 de 100 % en algunos casos. Estos resultados destacan el valor práctico de combinar sensores ópticos avanzados y redes neuronales para una evaluación rápida y no destructiva de la calidad y seguridad de la fruta. Este enfoque ofrece beneficios significativos para consumidores, productores y autoridades de seguridad alimentaria al mejorar la calidad del producto, reducir el desperdicio y garantizar el cumplimiento de las normativas de seguridad., [CA] Assegurar una bona qualitat interna e inocuitat en fruites és un repte important per a la indústria agroalimentària, ja que els mètodes tradicionals d'avaluació sovint són destructius i ineficients. Aquesta tesi doctoral ha tingut com a objectiu explorar l'ús de dues tècniques òptiques no destructives: la imatge hiperespectral (HSI) i l'espectroscòpia Raman, combinades amb models avançats de xarxes neuronals, per predir atributs de qualitat interna, acceptació del consumidor i seguretat alimentària en maduixes, caquis, prunes i pomes. La investigació va incloure quatre estudis principals. Primer, es van aplicar la imatge hiperespectral i les xarxes neuronals a 17 cultius de maduixa per predir paràmetres clau de qualitat com el contingut de sucre, l'acidesa i la textura, així com l'acceptació global del consumidor i la identificació dels cultius per prevenir el frau alimentari. En segon lloc, es van utilitzar models d'aprenentatge profund basats en autoencoders per classificar la imatge hiperespectral de caqui 'Rojo Brillante' en grups de textura rellevants per a la classificació a les centrals de confecció. Tercer, es van entrenar i comparar tres xarxes neuronals convolucionals (CNN) per detectar cops invisibles en prunes 'Presenta', amb l'objectiu de reduir les pèrdues postcollita. Quart, l'espectroscòpia Raman combinada amb CNN va permetre la detecció i identificació de múltiples compostos actius de plaguicides en mescles complexes de plaguicides habituals als pomerars. Els resultats van demostrar que la imatge hiperespectral i les xarxes neuronals van oferir prediccions precises per a la qualitat de la maduixa (TSS R² = 0,85, TA R² = 0,81), l'acceptació mitjana del consumidor per cultivar (R² = 0,78) i la discriminació dels cultius (F1 = 84 %). En caquis, els autoencoders van aconseguir una reducció del 99 % en la dimensionalitat de les dades i un 81,3 % d'exactitud en la classificació de la textura. Les CNN van detectar cops en fase inicial en les prunes amb una puntuació F1 del 90 % i es van identificar longituds d'ona clau. L'espectroscòpia Raman amb models d'aprenentatge profund va identificar la presencia de plaguicidas en mescles complexes amb un F1 de fins al 100% en alguns casos. Aquests resultats posen de manifest el valor pràctic de combinar tècniques òptiques avançades i xarxes neuronals per avaluar, de manera ràpida i no destructiva, la qualitat i seguretat dels fruits. Aquest enfocament ofereix beneficis significatius per als consumidors, productors i autoritats de seguretat alimentària, millorant la qualitat del producte, reduint el malbaratament i assegurant el compliment de les normes de seguretat., [EN] Ensuring high internal quality and safety in fruit products is a major challenge for the agri-food industry, as traditional assessment methods are often destructive and inefficient. This doctoral thesis aimed to explore the use of two non-destructive optical techniques: hyperspectral imaging (HSI) and Raman spectroscopy, combined with advanced neural network models, to predict internal quality attributes, consumer acceptance, and food safety in strawberries, persimmons, plums, and apples. The research involved four main studies. First, HSI and neural networks were applied to 17 strawberry cultivars to predict key quality parameters such as sugar content, acidity, and texture, as well as overall consumer acceptance and cultivar identification to prevent food fraud. Second, autoencoder-based deep learning models were used to classify 'Rojo Brillante' persimmon HSI into texture groups relevant for packing house sorting. Third, three convolutional neural networks (CNNs) were trained and compared to detect invisible mechanical bruises in 'Presenta' plums, aiming to reduce postharvest losses. Fourth, Raman spectroscopy combined with CNNs enabled the detection and identification of multiple pesticide active compounds in complex pesticide mixtures commonly used in apple orchards. The results demonstrated that HSI and neural networks provided accurate predictions for strawberry quality (TSS R² = 0.85, TA R² = 0.81), mean cultivar consumer acceptance (R² = 0.78), and cultivar discrimination (F1 = 84 %). In persimmons, autoencoders achieved a 99 % reduction in data dimensionality and 81.3 % accuracy in texture classification. CNNs detected early-stage bruises in plums with an F1 score of 90 % and key wavelengths were identified. Raman spectroscopy with deep learning models identified the presence of specific pesticides in complex mixtures with an F1 of up to 100 % in some cases. These findings highlight the practical value of combining advanced optical sensing and neural networks for rapid, non-destructive assessment of fruit quality and safety. This approach offers significant benefits for consumers, producers, and food safety authorities by improving product quality, reducing waste, and ensuring compliance with safety standards., Este trabajo ha sido parcialmente financiado por los proyectos MICIU AEI PID2023-150192OR-C31, C32 y C33, con el apoyo de fondos FEDER, los proyectos AEI TED2021-130117B-C31 y C33, el proyecto GVA-PROMETEO CIPROM/2021/014, así como por el proyecto PID2021-127975OR-C21. Asimismo, se ha contado con el apoyo de GVA-IVIA y fondos FEDER a través del proyecto 52204, y con financiación del proyecto Detection of synthetic chemical pesticides by Raman spectroscopy, respaldado por el Colruyt Group. Parte de este trabajo se enmarca en el proyecto PPS Smaakborging Groente en Fruit (Taste assurance of fruits and vegetables), financiado parcialmente por el Ministerio de Asuntos Económicos de los Países Bajos, Secretaría Top Sector Horticulture and Starting Materials. Salvador Castillo Gironés agradece la beca FPI-INIA (PRE2020-094491), parcialmente financiada por la Unión Europea (FSE), así como el apoyo de la COST Action CA19145 que hizo posible la estancia de investigación en KU Leuven. Sandra Munera agradece su contrato postdoctoral Juan de la Cierva - Formación (FJC2021-047786-I), cofinanciado por MCIN/AEI/10.13039/501100011033 y la Unión Europea NextGenerationEU/PRTR. Nariane Q. Vilhena agradece la financiación recibida a través de la beca PRE2018-085833 del Ministerio de Ciencia e Innovación de España. Remi Van Belleghem agradece la financiación otorgada por la Research Foundation-Flanders (FWO, Bruselas, Bélgica) mediante la beca Ph.D. Fellowship Strategic Basic Research (1S28522N). Los autores expresan su agradecimiento a Fresh Forward, Brookberries, Onethird y Bakker Barendrecht por la provisión de fruta, apoyo y supervisión técnica. También agradecen a Mieke Thoelen de Belorta y a Wim Wouters de Fruithandel Romain Wouters & Co. nv por su colaboración y apoyo logístico. Finalmente, se reconoce el respaldo del Walloon Agricultural Research Centre (CRA-W) por el suministro de muestras utilizadas en este estudio.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Artículo científico (JournalArticle)

Firmness classification in &apos, Rojo Brillante&apos, persimmon fruit after cold storage using hyperspectral imaging and deep learning with variational autoencoders

  • Castillo-Gironés, Salvador
  • González, Ángel
  • Martínez-Sober, Marcelino
  • Cubero-García, Sergio
  • Vilhena, Nariane Q.
  • Blanco-Álvarez, Iván
  • Gomez-Sanchis, Juan
  • Munera, S|||0000-0003-3064-1186
[EN] Firmness is one of the most important quality parameters of 'Rojo Brillante' persimmon fruit, and it can be significantly affected during cold storage. Industry requirements demand 40 N at harvest for storage and 20 N for commercialisation. This study evaluated the use of hyperspectral imaging to determine the firmness of persimmon stored under industrial conditions. To achieve this, fruit samples were stored for up to three months under various industrial conditions to produce a wide range of firmness levels. Hyperspectral images of each sample fruit were captured within the 420-1010 nm range, and the firmness was measured by a texturometer. An innovative methodology was applied, combining a convolutional variational autoencoder (VAE) for feature extraction from firmness curves and images and clustering analysis to identify firmness levels. The clustering analysis identified three distinct firmness categories (hard, medium and soft) corresponding to the requirements of the industry. Using six wavelengths, a support vector machine (SVM) classification model accurately assigned 81.3 % of fruit samples to the corresponding categories, thus paving the way for the utilisation of in-line detection using multispectral images. This approach demonstrates the potential of combining hyperspectral imaging and advanced machine learning techniques for non-destructive firmness prediction in 'Rojo Brillante' persimmon fruit quality assessment., This work was partially funded through the project AEI PID2023-150192OR-C31 and C33 with the support of FEDER funds, project AEI TED2021-130117B-C31 and C33, by the project GVA-PROMETEO CIPROM/2021/014 and by the project PID2021-127975OR-C21. Sal-vador Castillo thanks INIA for the FPI-INIA grant number PRE2020-094491, partially supported by European Union FSE funds. Nariane Q. Vilhena thanks the Ministerio de Ciencia e Innovacion of Spain for grant PRE2018-085833. Sandra Munera expresses her thanks for the post-doctoral contract Juan de la Cierva-Formacion (FJC2021-047786-I) co-funded by MCIN/AEI/10.13039/501100011033 and European Union NextGeneration EU/PRTR.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Artículo científico (JournalArticle)

Artificial Neural Networks in Agriculture, the core of artificial intelligence: What, When, and Why

  • Castillo-Gironés, Salvador
  • Martínez-Sober, Marcelino
  • Blasco, Jose
  • Cubero, Sergio
  • Gomez-Sanchis, Juan
  • Munera, S|||0000-0003-3064-1186
[EN] Artificial Neural Networks (ANNs) based models have emerged as a powerful tool for solving complex nonlinear problems in agriculture. These models simulate the human nervous system's structure, allowing them to learn hierarchical features from the data and solve nonlinear problems efficiently. Despite requiring a large amount of training data, ANNs with shallow architectures demonstrate superior performance in extracting relevant features and establishing accurate models, instilling confidence in their effectiveness compared to conventional machine learning methods. The versatility of ANNs enables their application in various agricultural domains, including precision agriculture, species classification, phenotyping, and food quality and safety assessment. ANNs combined with image analysis have proven valuable in disease detection, plant phenotyping, and fruit quality evaluation. The use of deep learning in agriculture has experienced exponential growth, as evident from the increasing number of publications in recent years. This article overviews recent advancements in applying ANNs in agriculture. It delves into the fundamental principles behind various types of agricultural data and ANN models, discussing their benefits and challenges. The article offers valuable insights into the proper use and functioning of each neural network, data processing for improved model outcomes, and the diverse applications of ANNs in the agricultural sector. It aims to equip readers with practical information on data utilisation, model selection based on data type, functionality, and current research applications., This work was partially funded through the projects MICIU AEI PID2023-150192OR-C31, C32 and C-33 with the support of FEDER funds, project AEI TED2021-130117B-C31 and C33, by the project GVA-PROMETEO CIPROM/2021/014. Salvador Castillo thanks INIA for the FPI-INIA grant number PRE2020-094491, partially supported by Euro-pean Union FSE funds. Sandra Munera expresses her thanks for the postdoctoral contract Juan de la Cierva-Formacion (FJC2021-047786-I) co-funded by MCIN/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Artículo científico (JournalArticle)

Automatic detection of pomegranate fruit affected by blackheart disease using X-ray imaging

  • Munera, S|||0000-0003-3064-1186
  • Rodriguez-Ortega, Alejandro|||0000-0003-4781-6368
  • Aleixos Borrás, María Nuria|||0000-0001-6051-3375
  • Cubero, Sergio
  • Blasco, José
[EN] Blackheart is one of the primary diseases affecting pomegranate fruit globally, caused by the fungus Alternaria. The damages are not visually detectable, as it is an internal disease that requires non-invasive technologies to provide information from inside the fruit to be detected. This study explored the ability of X-ray imaging to detect this infection in `Wonderful¿ pomegranate fruit. X-ray images of healthy and infected fruit at different levels were acquired and analysed. Texture features based on first-order statistics, the grey-level co-occurrence matrix (GLCM), and grey-level histograms with several resolutions were extracted from X-ray images and used to classify the fruit as healthy or infected through the random forest algorithm. The presence of the infection in three levels of severity was later assessed by destructive visual analysis by opening the samples in half. The highest accuracy models were obtained using all texture features and histograms with 256 bins. Compared to manual inspection, X-rays showed a clear advantage in detecting incipient infections (infected fruit at level 1), correctly identifying 93.3 % of infected fruits. In contrast, the manual inspection identified only 66.7 % of fruit, highlighting the limitations of early-stage detection., This work was partially funded by the projects GVA-PROMETEO CIPROM/2021/014, MICIU-AEI (PID2023-150192OR-C31 and C32) and GVA-IVIA 52204 with the EU through the European Regional Development Fund (ERDF) from GVA 2021 2027. Sandra Munera thanks the postdoctoral contract Juan de la Cierva-Formacion ¿ (FJC2021-047786-I) co-funded by MCIN/AEI/10.13039/ 501100011033 and European Union NextGenerationEU/PRTR.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Artículo científico (JournalArticle)

Raman spectroscopy for multi-label identification of common apple pesticide mixtures using CNNs and gradient-weighted class activation mapping

  • Castillo-Gironés, Salvador
  • Arnould, Quentin
  • Gomez-Sanchis, Juan
  • Blasco, Jose
  • Pigeon, Oliver
  • Baeten, Vicent
  • Fernandez-Pierna, Juan Antonio
[EN] Fruits on the market must be free of harmful pesticides, but residues often persist, posing a risk to public health. Detecting these residues efficiently is crucial. However, traditional methods are slow and lack real-time capability. Besides, the detection challenge becomes even more difficult when pesticides are used in mixtures, and Raman spectroscopy offers a promising real-time solution for identifying pesticide residues. This study investigates the detection of commonly used pesticide active compounds on apples in complex mixtures via Raman spectroscopy and a multilabel classification approach, addressing the challenge of identifying active compounds within complex mixtures. Models were also tested on independent, entirely different mixtures to assess their limitations and robustness. Unique spectral fingerprints were established for pure active compounds, which enabled differentiation in mixture analyses. Dimensionality reduction techniques (t-SNE and PCA) effectively distinguished pure compounds from mixtures. Three machine learning models: Partial Least Squares (PLS), Support Vector Machine (SVM), and a 1-D Convolutional Neural Network (1-D CNN) were trained to identify active compounds in mixtures. All models classified key compounds with high F1 scores, with Captan detected with F1 scores higher than 99 %, but some compounds, such as Folpet or Mancozeb, showed poor predictions. The 1-D CNN achieved the best performance with the lowest Hamming Loss, and Grad-CAM analysis confirmed it identified relevant spectral regions. This work demonstrates the potential of machine learning-enhanced Raman spectroscopy for effective pesticide monitoring, regulatory compliance, and food safety., This work is funded by GVA-IVIA and FEDER funds by project 52204, project PID2023-150192OR-C31, C32 and C33, project AEI TED2021-130117B-C31 and C33, by the project GVA-PROMETEO CIPROM/2021/014 and by the project PID2021-127975OR-C21. Salvador Castillo thanks INIA for the FPI-INIA grant number PRE2020-094491, partially supported by European Union FSE funds. We thank the Colruyt group for financial support during the project entitled 'Detection of synthetic chemical pesticides by Raman spectroscopy'. We also thank the Protection, Control Products and Residues Unit of the Walloon Agricultural Research Centre (CRA-W) for supplying the samples used in this work.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication Artículo científico (JournalArticle)

Advanced evaluation of strawberry quality, consumer preference, and cultivar discrimination through spectral imaging and neural networks

  • Castillo-Gironés, Salvador
  • Ruizendaal, Jos
  • Salas-Valderrama, Xiomara
  • Blasco, Jose
  • Polder, Gerrit
  • Munera, S|||0000-0003-3064-1186
[EN] Strawberries are among the most popular fruits, and meeting the rising demand for high-quality, flavorful varieties requires understanding consumer preferences. Accurately predicting these preferences, assessing quality, and preventing food fraud are crucial for breeders and sellers. This helps breeders develop superior cultivars and enables sellers to sort and market strawberries by taste and quality. This study explores the prediction of the quality and the acceptance of Dutch consumers of seventeen strawberry cultivars and their discrimination using VIS-NIR spectral imaging with a spectral range between 400 and 1000 nm and Artificial Neural Networks (ANNs), which was not done before. A total of 3564 samples were utilized. Three algorithms: Support Vector Machine, XGBoost, and a Multilayer Perceptron (MLP), were evaluated to predict quality parameters, consumer acceptance, and cultivar discrimination. MLP models showed the highest accuracy, with R2 values of 0.85 for total soluble solids, 0.81 for titratable acidity, 0.76 for bite, and 0.78 for overall consumer acceptance. For cultivar discrimination, the MLP model achieved an F1 score of 0.84. These findings highlight the potential of ANNs in enhancing product quality assessment, preventing food fraud, and aligning products with consumer preferences in the food industry., This work is part of the project PPS Smaakborging Groente en Fruit (Taste assurance of fruits and vegetables) , partly funded by the Ministry of Economic Affairs, Secretariat Top Sector Horticulture and Starting Materials. The authors thank Fresh Forward, Brookberries, Onethird and Bakker Barendrecht for providing the fruit, support and technical supervision. Also, of MICIU AEI PID2023-150192OR-C31, C32 and C-33 with the support of FEDER funds and GVA-PROMETEO CIPROM/2021/014. Salvador Castillo Girones thanks INIA for the FPI-INIA grant number PRE2020-094491, co-funded by EU EFS funds. Sandra Munera thanks the postdoctoral contract Juan de la Cierva-Formacion (FJC2021-047786-I) co-funded by MCIN/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR.
ReDivia Repositorio Digital del Instituto Valenciano de Investigaciones Agrarias
Publication Artículo científico (JournalArticle)

Automatic detection of pomegranate fruit affected by blackheart disease using X-ray imaging

  • Munera, Sandra
  • Rodríguez-Ortega, Alejandro
  • Cubero, Sergio
  • Aleixos, Nuria
  • Blasco, José
Blackheart is one of the primary diseases affecting pomegranate fruit globally, caused by the fungus Alternaria. The damages are not visually detectable, as it is an internal disease that requires non-invasive technologies to provide information from inside the fruit to be detected. This study explored the ability of X-ray imaging to detect this infection in ‘Wonderful’ pomegranate fruit. X-ray images of healthy and infected fruit at different levels were acquired and analysed. Texture features based on first-order statistics, the grey-level co-occurrence matrix (GLCM), and grey-level histograms with several resolutions were extracted from X-ray images and used to classify the fruit as healthy or infected through the random forest algorithm. The presence of the infection in three levels of severity was later assessed by destructive visual analysis by opening the samples in half. The highest accuracy models were obtained using all texture features and histograms with 256 bins. Compared to manual inspection, X-rays showed a clear advantage in detecting incipient infections (infected fruit at level 1), correctly identifying 93.3 % of infected fruits. In contrast, the manual inspection identified only 66.7 % of fruit, highlighting the limitations of early-stage detection.