<p dir="ltr">Among the major row crops in the United States, corn and soybeans stand out due to their high nutritional value and economic importance. Achieving optimal yields is restrained by the challenge of fertilizer management. Many fields experience yield losses due to insufficient mineral nutrients like nitrogen (N), while excessive fertilization raises costs and environmental risks. The critical issue is the accurate determination of fertilizer quantity and timing, underscoring the need for precise, early-stage diagnostics. Emerging high-throughput plant phenotyping techniques, notably hyperspectral imaging (HSI), have been increasingly utilized to identify plant’s responses to abiotic or biotic stresses. Varieties of HSI systems have been developed, such as airborne imaging systems and indoor imaging stations. However, most of the current HSI systems’ signal quality is often compromised by various environmental factors. To address the issue, a handheld hyperspectral imager known as LeafSpec was recently developed at Purdue University and represents a breakthrough with its ability to scan corn or soybean leaves at exceptional spatial and spectral resolutions, improving plant phenotyping quality at reduced costs. Most of the current HSI data processing methods focus on spectral features but rarely consider spatially distributed information. Thus, the objective of this work was to develop a methodology utilizing spatial-spectral features for accurate and reliable diagnostics of crop N nutrient stress. The key innovations include the designing of spatial-spectral features based on the leaf venation structures and the feature mining method for predicting the plant nitrogen condition. First, a novel analysis method called the Natural Leaf Coordinate System (NLCS) was developed to reallocate leaf pixels and innovate the nutrient stress analysis using pixels’ relative locations to the venation structure. A new nitrogen prediction index for soybean plants called NLCS-N was developed, outperforming the conventional averaged vegetation index (Avg. NDVI) in distinguishing healthy plants from nitrogen-stressed plants with higher t-test p-values and predicting the plant nitrogen concentration (PNC) with higher R-squared values. In one of the test cases, the p-values and R-squared values were improved, respectively, from 2.1×10^-3 to 6.92×10^-12 and from 0.314 to 0.565 by Avg. NDVI and NLCS-N. Second, a corn leaf venation segmentation algorithm was developed to separate the venation structure from a corn leaf LeafSpec image, which was further used to generate 3930 spatial-spectral (S-S) features. While the S-S features could be the input variable to build a PNC prediction model, a feature selection mechanism was developed to improve the models’ accuracy in terms of reduced cross-validation errors. In one of the test cases, the cross-validation root mean squared errors were reduced compared with the leaf mean spectra from 0.273 to 0.127 using the selected features. Third, several novel spatial-spectral indexes for corn leaves were developed based on the color distributions at the venation level. The top-performing indexes were selected through a ranking system based on Cohen’s d values and the R-squared values, resulting in a best-performing S-S N prediction index with 0.861 R-squared values for predicting the corn PNC in a field assay. The discussion sections provided insights into how a robust PNC prediction index could be developed and related to plant science. The methodologies outlined offer a framework for broader applications in spatial-spectral analysis using leaf-level hyperspectral imagery, serving as a guide for scientists and researchers in customizing their future studies within this field.</p>

With the evolution of GPU performance, the interest of using AI for all kinds of purposes has risen. Companies today put a great amount of resources to find new ways of using AI to increase the value of their products or automating processes. An area in the wood industry where AI is widely used and studied is in defect detection. In this thesis, the combination of using AI and hyperspectral images is studied and evaluated in the case of segmenting defects in hardwood with a U- Net network structure. The performance is compared to another known method usually used when dealing with high-dimensional data: PLS-DA. This thesis also compares the use of RGB image data in combination with AI, to further analyze the usefulness that the hyperspectral data provide. The results showed signs of improvement when using hyperspectral images com- pared to RGB images when detecting blue stain and red heartwood defects. De- tection of the defects rot and knots did however show no sign of improvements. Due to the annotations being more accurate in the RGB data, the results from the hyperspectral data-fed networks would suggest that blue stain and red heartwood could be of interest regarding further investigation. Computational performance is shown to vary across the different reduction meth- ods, and the results from this thesis provides some insight that might aid in the reasoning regarding how to choose an appropriate reduction method.

Hyper-Raman-Streuung folgt anderen Symmetrieauswahlregeln als Raman-Streuung und profitiert als nicht-linearer Zweiphotonenprozess noch mehr von verstärkten elektromagnetischen Feldern an der Oberfläche plasmonischer Nanostrukturen. Damit könnte die oberflächenverstärkte Hyper-Raman-Streuung (SEHRS) praktische Bedeutung in der Spektroskopie erlangen. Durch die Kombination von SEHRS und oberflächenverstärkter Raman-Streuung (SERS) können komplementäre Strukturinformationen erhalten werden. Diese eignen sich aufgrund der Lokalisierung der Verstärkung auf die unmittelbare Umgebung der Nanostrukturen besonders für die Charakterisierung der Wechselwirkung zwischen Molekülen und Metalloberflächen. Ziel dieser Arbeit war es, ein tieferes Verständnis des SEHRS-Effekts zu erlangen und dessen Anwendbarkeit für analytische Fragestellungen einzuschätzen. Dazu wurden SEHRS-Experimente mit Anregung bei 1064 nm und SERS-Experimente mit Anregung bei derselben Wellenlänge sowie mit Anregung bei 532 nm - für eine Detektion von SEHRS und SERS im gleichen Spektralbereich - durchgeführt. Als Beispiel für nicht-resonante Anregung wurden die vom pH-Wert abhängigen SEHRS- und SERS-Spektren von para-Mercaptobenzoesäure untersucht. Mit diesen Spektren wurde die Wechselwirkung verschiedener Silbernanostrukturen mit den Molekülen charakterisiert. Anhand von beta-Carotin wurden Einflüsse von Resonanzverstärkung im SEHRS-Experiment durch die gleichzeitige Anregung eines molekularen elektronischen Übergangs untersucht. Dabei wurde durch eine Thiolfunktionalisierung des Carotins eine intensivere Wechselwirkung mit der Silberoberfläche erzielt, sodass nicht nur resonante SEHRS- und SERS-Spektren, sondern auch nicht-resonante SERS-Spektren von Carotin erhalten werden konnten. Die Anwendbarkeit von SEHRS für hyperspektrale Kartierung in Verbindung mit Mikrospektroskopie wurde durch die Untersuchung von Verteilungen verschiedener Farbstoffe auf strukturierten plasmonischen Oberflächen demonstriert. / Hyper-Raman scattering follows different symmetry selection rules than Raman scattering and, as a non-linear two-photon process, profits even more than Raman scattering from enhanced electromagnetic fields at the surface of plasmonic nanostructures. Surface-enhanced hyper-Raman scattering (SEHRS) could thus gain practical importance for spectroscopy. The combination of SEHRS and surface-enhanced Raman scattering (SERS) offers complementary structural information. Specifically, due to the localization of the enhancement to the close proximity of the nanostructures, this information can be utilized for the characterization of the interaction between molecules and metal surfaces. The aim of this work was to increase the understanding of the SEHRS effect and to assess its applicability to answer analytical questions. For that purpose, SEHRS experiments with excitation at 1064 nm and SERS experiments with excitation at the same wavelength, as well as with excitation at 532 nm - to detect SEHRS and SERS in the same spectral region - were conducted. As an example for non-resonant excitation, pH-dependent SEHRS and SERS spectra of para-mercaptobenzoic acid were examined. Based on these spectra, the interaction of different silver nanostructures with the molecules was characterized. beta-Carotene was used to study the influence of resonance enhancement by the excitation of a molecular electronic transition during SEHRS experiments. By the thiol-functionalization of carotene, a more intense interaction with the silver surface was achieved, which enables to obtain not only resonant SEHRS and SERS but also non-resonant SERS spectra of carotene. Hyperspectral SEHRS imaging in combination with microspectroscopy was demonstrated by analyzing the distribution of different dyes on structured plasmonic surfaces.

Total soluble solids (TSS) is a reliable indicator of melon eating quality, with a minimum standard of 10% recommended. The state of Australian melon production with respect to this quality criterion was considered within seasons, between growing districts and over seasons. It was concluded that improvement in agronomic practice and varietal selection is required to produce sweeter melons. The scientific literature addressing melon physiology and agronomy was summarised, as a background to the work that is required to improve melon production practices in Australia. The effect of source sink manipulation was assessed for commercially grown and glasshouse grown melon plants. The timing of fruit thinning, pollination scheduling, the application of a growth inhibitor and source biomass removal were assessed in relation to fruit growth and sugar accumulation. Results are interpreted against a model in which fruit rapidly increase in weight until about two weeks before harvest, with sugar accumulation continuing as fruit growth ceases. Thus treatment response is very dependant on timing of application. For example, fruit thinning at 25 days before harvest resulted in further fruit set and increased fruit weight but did not impact on fruit TSS (at 9.8%, control 9.3%), while thinning at 5 days before harvest resulted in a significant (Pless than 0.05) increase in fruit TSS (to 10.8%, control 9.3%) and no increase in fruit weight or number. A cost/ benefit analysis is presented, allowing an estimation of the increase in sale price required to sustain the implementation of fruit thinning. The effect of irrigation scheduling was also considered with respect to increasing melon yield and quality. To date, recommended practice has been to cause an irrigation deficit close to fruit harvest, with the intent of 'drying out' or 'stressing' the plant, to 'bring on' maturity and increase sugar accumulation. Irrigation trials showed that keeping plants stress-free close to harvest and during harvest, facilitated the production of sweeter fruit. The maintenance of a TSS grade standard using either batch based (destructive) sampling or (non-invasive) grading of individual fruit is discussed. On-line grading of individual fruit is possible using near infrared spectroscopy (NIR), but the applicability of the technique to melons has received little published attention. Tissue sampling strategy was optimised, in relation to the optical geometry used (in commercial operation in Australia), both in terms of the diameter and depth of sampled tissue. NIR calibration model performance was superior when based on the TSS of outer, rather than inner mesocarp tissue. However the linear relationship between outer and middle tissue TSS was strong (r2 = 0.8) in immature fruit, though less related in maturing fruit (r2 = 0.5). The effect of fruit storage (maturation/senescence) on calibration model performance was assessed. There was a negligible effect of fruit cold storage on calibration performance. Currently, the agronomist lacks a cost-effective tool to rapidly assess fruit TSS in the field. Design parameters for such a tool were established, and several optical front ends compared for rapid, though invasive, analysis. Further, for visualisation of the spatial distribution of tissue TSS within a melon fruit, a two-dimensional, or hyper-spectral NIR imaging system based on a low cost 8-bit charge coupled device (CCD) camera and filter arrangement, was designed and characterised.

<p>Hyperspectral reflectance data are utilised in many applications, where measured data are processed and converted into physical, chemical and/or biological properties of the target objects and/or processes being studied. It has been proven that crop reflectance data can be used to detect, characterise and quantify disease severity and plant density.</p><p>In this thesis, various methods were proposed and used for detection, characterisation and quantification of disease severity and plant density utilising data acquired by hand-held spectrometers. Following this direction, hyperspectral images provide both spatial and spectral information opening for more efficient analysis.</p><p>Hence, in this thesis, various surface water quality parameters of inland waters have been monitored using hyperspectral images acquired by airborne systems. After processing the images to obtain ground reflectance data, the analysis was performed using similar methods to those of the previous case. Hence, these methods may also find application in future satellite based hyperspectral imaging systems.</p><p>However, the large size of these images raises the need for efficient data reduction. Self organising and learning neural networks, that can follow and preserve the topology of the data, have been shown to be efficient for data reduction. More advanced variants of these neural networks, referred to as the weighted neural networks (WNN), were proposed in this thesis, such as the weighted incremental neural network (WINN), which can be used for efficient reduction, mapping and clustering of large high-dimensional data sets, such as hyperspectral images.</p><p>Finally, the analysis can be reversed to generate spectra from simpler measurements using multiple colour-filter mosaics, as suggested in the thesis. The acquired instantaneous single image, including the mosaic effects, is demosaicked to generate a multi-band image that can finally be transformed into a hyperspectral image.</p>

Hyperspectral reflectance data are utilised in many applications, where measured data are processed and converted into physical, chemical and/or biological properties of the target objects and/or processes being studied. It has been proven that crop reflectance data can be used to detect, characterise and quantify disease severity and plant density. In this thesis, various methods were proposed and used for detection, characterisation and quantification of disease severity and plant density utilising data acquired by hand-held spectrometers. Following this direction, hyperspectral images provide both spatial and spectral information opening for more efficient analysis. Hence, in this thesis, various surface water quality parameters of inland waters have been monitored using hyperspectral images acquired by airborne systems. After processing the images to obtain ground reflectance data, the analysis was performed using similar methods to those of the previous case. Hence, these methods may also find application in future satellite based hyperspectral imaging systems. However, the large size of these images raises the need for efficient data reduction. Self organising and learning neural networks, that can follow and preserve the topology of the data, have been shown to be efficient for data reduction. More advanced variants of these neural networks, referred to as the weighted neural networks (WNN), were proposed in this thesis, such as the weighted incremental neural network (WINN), which can be used for efficient reduction, mapping and clustering of large high-dimensional data sets, such as hyperspectral images. Finally, the analysis can be reversed to generate spectra from simpler measurements using multiple colour-filter mosaics, as suggested in the thesis. The acquired instantaneous single image, including the mosaic effects, is demosaicked to generate a multi-band image that can finally be transformed into a hyperspectral image.

Tesis por compendio / [ES] Las técnicas rápidas, no destructivas y libres de químicos tienen una demanda creciente en muchos campos de la industria. Las técnicas de espectroscopia de infrarrojo cercano (NIRS) y imágenes hiperespectrales NIR (NIR-HSI) han mostrado un gran potencial para determinar los parámetros de calidad de los alimentos, autenticar productos alimenticios, detectar el fraude, entre otras. En la NIRS, las medidas se toman en puntos específicos, detectando solo una pequeña porción; en la NIR-HSI, la información espectral y espacial se combinan, lo que la convierte en una opción adecuada para muchos productos alimenticios, ya que son matrices muy heterogéneas. Por lo tanto, este estudio tuvo como objetivo revisar la aplicación de NIRS (dispersivos), NIR de Transformada de Fourier (FT) y HSI en la evaluación de los parámetros de calidad de harina de trigo y productos a base de trigo, así como para la autenticación y determinación de la composición de estos productos. Además, este trabajo tuvo como objetivo identificar y clasificar diferentes tipos de muestras de fibra agregadas a la semolina y pasta producidas por estas formulaciones, y monitorear el proceso de cocción de esta pasta enriquecida en fibra mediante técnicas espectrales. Además, se objetivó aplicar HSI a otro producto en polvo, por lo que se cuantificó el contenido de pectina en las cáscaras de naranja. Primero, se adquirieron espectros NIR para comparar la precisión en la clasificación de muestras enriquecidas con fibra, para cuantificar la cantidad de estas fibras y verificar su distribución en muestras de semolina. Para la clasificación se utilizaron el Análisis de Componentes Principales (PCA) y el Soft Independent Modelling of Class Analogy (SIMCA). Los modelos de regresión de mínimos cuadrados parciales (PLSR) aplicados a espectros NIR-HSI mostraron R²P entre 0,85 y 0,98 y RMSEP entre 0,5 y 1, y los modelos se utilizaron para construir los mapas químicos para verificar la distribución de fibra en las superficies de las muestras. Además, se probó el NIR-HSI junto con Multivariate Curve Resolution-Alternating Least Squares (MCR-ALS) para investigar la capacidad de evaluación, resolución y cuantificación de la distribución de fibra en la pasta. Los resultados mostraron R²P entre 0.28 y 0.89,% de falta de ajuste (LOF) <6%, varianza explicada sobre 99% y similitud entre espectros puros y recuperados sobre 96% y 98%. Además, se probó VIS/NIR-HSI en el modo de transmisión como una alternativa objetiva para la clasificación de muestras de pasta según el tiempo de cocción. El análisis discriminante lineal (LDA) mostró valores de sensibilidad y especificidad entre 0,14-1,00 y 0,51-1,00, respectivamente, y una tasa de ausencia de error (NER) superior a 0,62. El análisis discriminante de mínimos cuadrados parciales (PLSDA) mostró valores de sensibilidad y especificidad entre 0,67-1,00 y 0,10-1,00, respectivamente, y NER superiores a 0,80. Los resultados de este trabajo mostraron que la técnica NIR-HSI se puede utilizar para la identificación y cuantificación de la fibra agregada a la semolina. Además, NIR-HSI y MCR-ALS pueden identificar la fibra en la pasta. La HSI en el modo de transmisión demostró ser una técnica adecuada como alternativa objetiva para la clasificación de muestras de pasta según el tiempo de cocción como una forma de automatizar la determinación de los atributos de la pasta. La determinación del contenido de pectina en cáscaras de naranja se investigó usando NIR-HSI. LDA mostró mejores resultados de discriminación considerando tres grupos: bajo (0-5%), intermedio (10-40%) y alto (50-100%) contenido. Los modelos PLSR basados en espectros completos mostraron mayor precisión (R2> 0,93, RMSEP entre 6,50 y 9,16% de pectina) que los basados en pocas longitudes de onda seleccionadas (R2 entre 0,92 y 0,94, RMSEP entre 8,03 y 9,73% de pectina). Los resultados demuestran el potencial de NIR-HSI para cuantificar el contenido de pectina en las cáscaras de naranja, proporcionando una técnica valiosa para los productores de naranja y las industrias de procesamiento. / [CA] Les tècniques ràpides, no destructives i lliures de químics tenen una demanda creixent en molts camps de la indústria. Les tècniques d'espectroscopia d'infraroig proper (NIRS) i d'imatges hiperespectrals NIR (NIR-HSI) han demostrat tindre un gran potencial per a determinar paràmetres de qualitat d'aliments, autenticar productes alimentaris, detectar frau entre altres aplicacions. Mentre que en la NIRS proper les mesures es prenen en punts específics de la mostra i es detecta una porció menuda, en la HSI es combina informació espectral i espacial de tal manera que és una opció adient per a molts tipus de productes alimentaris, ja que són matrius molt heterogènies. Per tant, este estudi va tindre com objectiu revisar tota l'aplicació de NIRS (dispersius), NIR de Transformada de Fourier (FT) i HSI en l'avaluació dels paràmetres de qualitat de la farina de blat i els productes a base de blat, així com per a l'autenticació i determinació de la composició d'estos productes. A més a més, este estudi va tindre com objectiu identificar i classificar diferents tipus de mostres de fibra afegides a la semolina i pasta produïdes per formulació de fibra i semolina, i monitorar mitjançant tècniques espectrals el procés de cocció d'aquesta pasta enriquida amb fibra. A més, este treball va tindre com objectiu aplicar HSI a un altre producte en pols, de tal manera que es va quantificar el contingut de pectina en les corfes de taronja. Primer, es van adquirir espectres NIR per comparar la precisió en la classificació de mostres enriquides amb fibra, per quantificar estes fibres i verificar la seua distribució en mostres de sèmola. Per a la classificació es van emprar l'Anàlisi de Components Principals (PCA) i el SIMCA (Soft Independent Modelling of Class Analogy). Els models de regressió de mínims quadrats parcials (PLSR) aplicats a espectres NIR-HSI mostraren R²P entre 0,85 i 0,98 i RMSEP entre 0,5 i 1% de contingut de fibra, i els models s'utilitzaren per construir els mapes químics per verificar la distribució de fibra en les superficies de les mostres. Així mateix, es va provar NIR-HSI amb Multivariate Curve Resolution-Alternating Least Square (MCR-ALS) per a investigar la capacitat d'avaluació, resolució i quantificació de la distribució de fibra en la pasta enriquida. Els resultats mostraren un R²P entre 0,28 i 0,89%, lack of fit (LOF)<6%, variància explicada sobre 99% i similitud entre espectres purs i recuperats sobre 96% i 98%. D'altra part, es va provar VIS/NIR-HSI en el mode de transmissió com una alternativa objectiva per a la classificació de mostres de pasta segons el temps de cocció. L'anàlisi discriminant lineal (LDA) va mostrar valors de sensibilitat i especificitat entre 0,14-1,00 i 0,51-1,00 respectivament, i una taxa d'absència d'error (NER) superior a 0,62. L'anàlisi discriminant de mínims quadrats parcials (PLSDA) va mostrar valors de sensibilitat i especificitat entre 0,67-1,00 i 0,10-1,00 respectivament, i NER superiors a 0,80. Els resultats d'este treball mostraren que la tècnica NIR-HSI es pot emprar per a la identificació i quantificació de la fibra afegida a la semolina. A més a més, NIR-HSI i MCR-ALS poden identificar la fibra en la pasta. La HSI en mode de transmissió va demostrar ser una tècnica adient com a alternativa objectiva per a la classificació de mostres de pasta segons el temps de cocció com forma d'automatitzar la determinació dels atributs de la pasta. La determinació del contingut de pectina en corfa de taronja es va investigar emprant NIR-HSI. LDA va mostrar millors resultats de discriminació considerant tres grups: baix (0-5%), intermedi (10-40%) i alt (50-100%). Els models PLSR basats en espectres complets van mostrar major precisió (R2> 0,93, RMSEP entre 6,50 i 9,16% de pectina) que els basats en longituds d’ona seleccionades (R2 entre 0,92 i 0,94, RMSEP entre 8,03 i 9,73% de pectina). Els resultats demostren el potencial de NIR-HSI per a quantificar el contingut de pectina en corfa de taronja i proporcionen una tècnica valuosa per als productors de taronja i les indústries de processament. / [EN] Fast, non-destructive and chemical-free techniques are in increasing demand in many fields of the industry. Near-infrared spectroscopy (NIRS) and NIR hyperspectral imaging (NIR-HSI) techniques have shown great potential in determining food quality parameters, authenticating food products, detecting food fraud, among many other applications. While in near infrared spectroscopy, the measurements are taken at specific points on the sample, detecting only a small portion; in hyperspectral imaging, spectral and spatial information are combined, making it a suitable choice for many food products, since they are very heterogeneous matrices. Therefore, this study aimed to review all the application of (dispersive) NIRS, Fourier Transform (FT) NIR, and HSI in assessing wheat flour and wheat-based products quality parameters, as well for the authentication and determination of composition of these products. Moreover, this work aimed to identify and classify different types of fibre samples added to the semolina and pasta produced by semolina-fibre formulations, and to monitor the cooking process of this fibre-enriched pasta by spectral techniques. In addition, this work had the aim of applying HSI to other powdered product, so the pectin content in orange peels was quantified. First, NIR spectra were acquired to compare the accuracy in the classification of fibre-enriched samples, to quantify the amount of these fibres and verify their distribution on semolina samples. Principal Component Analysis (PCA) and Soft Independent Modelling of Class Analogy (SIMCA) were used for classification. Partial Least Squares Regression (PLSR) models applied to NIR-HSI spectra showed R2P between 0.85 and 0.98, and RMSEP between 0.5 and 1% of fibre content, and the models were used to construct the chemical maps to check the fibre distribution on the samples surface. Moreover, NIR-HSI together with Multivariate Curve Resolution-Alternating Least Squares (MCR-ALS), was tested to investigate the ability for the evaluation, resolution and quantification of fibre distribution in enriched pasta. Results showed coefficient of determination of validation (R²V) between 0.28 and 0.89, % of lack of fit (LOF) <6%, variance explained over 99%, and similarity between pure and recovered spectra over 96% and 98% in models using pure flour and control as initial estimates, respectively. In addition, VIS/NIR-HSI in the transmission mode was tested as an objective alternative for the classification of pasta samples according to cooking time as way of automating the determination of pasta attributes. Linear Discriminant Analysis (LDA) showed values of sensitivity and specificity between 0.14-1.00 and 0.51-1.00, respectively, and non-error rate (NER) over 0.62. Partial Least Square Discriminant Analysis (PLSDA) showed values of sensitivity and specificity between 0.67 - 1.00 and 0.10-1.00, respectively, and NER over 0.80. The results of the first part of this work showed that NIR-HSI technique can be used for the identification and quantification of fibre added to semolina. Additionally, NIR-HSI and MCR-ALS are able to identify fibre in pasta. Hyperspectral imaging in the transmission mode demonstrated to be a suitable technique as an objective alternative for the classification of pasta samples according to the cooking time as a way of automating the determination of pasta attributes. Determination of pectin content in orange peels was investigated using NIR-HSI. LDA showed better discrimination results considering three groups:low(0-5%), intermediate(10-40%) and high(50-100%) pectin content. PLSR models based on full spectra showed higher precision (R²>0.93, RMSEP between 6.50 and 9.16% of pectin) than those based on few selected wavelengths (R² between 0.92 and 0.94, RMSEP between 8.03 and 9.73%). The results demonstrate the potential of NIR-HSI to quantify pectin content in orange peels, providing a valuable technique for orange producers and processing industries. / This work was supported by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior- Brasil (CAPES) [Finance Code 001]; São Paulo Research Foundation (FAPESP) [grant numbers 2015/24351-2, 2017/17628-3, 2019/06842- 0]; and by projects AEI PID2019-107347RR-C31 and PID2019-107347RR-C32, and the European Union through the European Regional Development Fund (ERDF) of the Generalitat Valenciana 2014-2020. The authors would like to thank Nutrassim Food Ingredients company for the donation of the fibre samples, the support provided by Enrique Aguilar María, Carlos Alberto Velasquez Hernández, Diego Hernández Catalán, Carlos Ruiz Catalá and Andrés Estuardo Prieto López during system installation, experimental analysis and data acquisition. / Teixeira Badaró, A. (2021). Nir Spectral Techniques and Chemometrics Applied to Food Processing [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/178758 / TESIS / Compendio

No description available.

<p>  </p> <p>Hyperspectral Imaging is one of the most popular technologies in plant phenotyping due to its ability to predict the plant physiological features such as yield biomass, leaf moisture, and nitrogen content accurately, non-destructively, and efficiently. Various kinds of hyperspectral imaging systems have been developed in the past years for both greenhouse and field phenotyping activities. Developing the plant physiological prediction model such as relative water content (RWC) using hyperspectral imaging data requires the adoption of machine learning-based calibration techniques. Convolutional neural networks (CNNs) have been known to automatically extract the features from the raw data which can lead to highly accurate physiological prediction models. Once a reliable prediction model has been developed, sharing that model across multiple hyperspectral imaging systems is very desirable since collecting the large number of ground truth labels for predictive model development is expensive and tedious. However, there are always significant differences in imaging sensors, imaging, and environmental conditions between different hyperspectral imaging facilities, which makes it difficult to share plant features prediction models. Calibration transfer between the imaging systems is critically important. In this thesis, two approaches were taken to address the calibration transfer from the greenhouse to the field. First, direct standardization (DS), piecewise direct standardization (PDS), double window piecewise direct standardization (DPDS) and spectral space transfer (SST) were used for standardizing the spectral reflectance to minimize the artifacts and spectral differences between different greenhouse imaging systems. A linear transformation matrix estimated using SST based on a small set of plant samples imaged in two facilities reduced the root mean square error (RMSE) for maize physiological feature prediction significantly, i.e., from 10.64% to 2.42% for RWC and from 1.84% to 0.11% for nitrogen content. Second, common latent space features between two greenhouses or a greenhouse and field imaging system were extracted in an unsupervised fashion. Two different models based on deep adversarial domain adaptation are trained, evaluated, and tested. In contrast to linear standardization approaches developed using the same plant samples imaged in two greenhouse facilities, the domain adaptation extracted non-linear features common between spectra of different imaging systems. Results showed that transferred RWC models reduced the RMSE by up to 45.9% for the greenhouse calibration transfer and 12.4% for a greenhouse to field transfer. The plot scale evaluation of the transferred RWC model showed no significant difference between the measurements and predictions. The methods developed and reported in this study can be used to recover the performance plummeted due to the spectral differences caused by the new phenotyping system and to share the knowledge among plant phenotyping researchers and scientists.</p>

Stains and stained incipient lesions can be challenging to differentiate with established clinical tools. New diagnostic techniques are required for improved distinction to enable early noninvasive treatment. This in vitro study evaluates the performance of artificial intelligence (AI)-based classification of hyperspectral imaging data for early occlusal lesion detection and differentiation from stains. Sixty-five extracted permanent human maxillary and mandibular bicuspids and molars (International Caries Detection and Assessment System [ICDAS] II 0–4) were imaged with a hyperspectral camera (Diaspective Vision TIVITA® Tissue, Diaspective Vision, Pepelow, Germany) at a distance of 350 mm, acquiring spatial and spectral information in the wavelength range 505–1000 nm; 650 fissural spectra were used to train classification algorithms (models) for automated distinction between stained but sound enamel and stained lesions. Stratified 10-fold cross-validation was used. The model with the highest classification performance, a fine k-nearest neighbor classification algorithm, was used to classify five additional tooth fissural areas. Polarization microscopy of ground sections served as reference. Compared to stained lesions, stained intact enamel showed higher reflectance in the wavelength range 525–710 nm but lower reflectance in the wavelength range 710–1000 nm. A fine k-nearest neighbor classification algorithm achieved the highest performance with a Matthews correlation coefficient (MCC) of 0.75, a sensitivity of 0.95 and a specificity of 0.80 when distinguishing between intact stained and stained lesion spectra. The superposition of color-coded classification results on further tooth occlusal projections enabled qualitative assessment of the entire fissure’s enamel health. AI-based evaluation of hyperspectral images is highly promising as a complementary method to visual and radiographic examination for early occlusal lesion detection.