Deep Learning Regression Model Research Articles

Nighttime Street View Imagery: A new perspective for sensing urban lighting landscape

Urban lighting reflects nocturnal activities and it is traditionally observed using Nighttime Lights (NTL) satellite imagery. Few studies systematically measure the nightscape from a human perspective. This study brings a new paradigm — urban lighting sensing via Nighttime Street View Imagery (SVI). The paradigm draws on the accomplishments of (daytime) SVI and gives attention to its ignored nighttime counterpart. We put forward this idea by manually collecting 2,831 nighttime SVIs across various urban functional areas in Singapore. We investigated their values by developing a use case for clustering nighttime lighting patterns. To mitigate the scarcity of nighttime SVI, deep learning regression models were trained to predict nighttime brightness based on corresponding daytime SVIs obtained from widely available sources. The results were compared with brightness data derived from satellite imagery, to affirm the novelty and uniqueness of nighttime SVI. As a result, there are 7 lighting patterns within the collected nighttime SVI, distinct in lighted spot features and total brightness. The identified patterns effectively characterize different urban function scenarios. The best trained brightness prediction model performs well in revealing the city-scale lighting landscape. The SVI-predicted brightness shows a distribution similar to the brightness from satellite imagery and complements it in urban areas with complex vertical lighting structures. This study demonstrates the potential of nighttime SVI as a valuable data source for mapping urban lighting and activities, offering advantages over satellite data. The proposed paradigm contributes significantly to cross-modal information mining in urban studies and has potential applications in scenarios such as light pollution mitigation and crime prevention.

Deep Learning-Based Fatigue Strength Prediction for Ferrous Alloy

As industrial development drives the increasing demand for steel, accurate estimation of the material’s fatigue strength has become crucial. Fatigue strength, a critical mechanical property of steel, is a primary factor in component failure within engineering applications. Traditional fatigue testing is both costly and time-consuming, and fatigue failure can lead to severe consequences. Therefore, the need to develop faster and more efficient methods for predicting fatigue strength is evident. In this paper, a fatigue strength dataset was established, incorporating data on material element composition, physical properties, and mechanical performance parameters that influence fatigue strength. A machine learning regression model was then applied to facilitate rapid and efficient fatigue strength prediction of ferrous alloys. Twenty characteristic parameters, selected for their practical relevance in engineering applications, were used as input variables, with fatigue strength as the output. Multiple algorithms were trained on the dataset, and a deep learning regression model was employed for the prediction of fatigue strength. The performance of the models was evaluated using metrics such as MAE, RMSE, R2, and MAPE. The results demonstrated the superiority of the proposed models and the effectiveness of the applied methodologies.

An explainable AI for green hydrogen production: A deep learning regression model

Currently, hydrogen generation is considered a crucial aspect of sustainable energy production. This paper disscusses the hypothesis that hydrogen generation occurs during the operation of the membrane electrode assembly and is an acceptable and eco-friendly method for producing hydrogen without carbon dioxide. This paper considers the hypothesis by simulating hydrogen generation in fuel cells that rely on fossil fuels using proton fuel cells. The Optimized Hydrogen Generation-based Regression (OHGR) model is based on a deep learning architecture. The architecture is based on stacking multiple neural networks with crossover connections that feed information through a pool of activation functions. The OHGR predicts the amount of hydrogen generated in 1 mA cm−2 and can exceed 10–20 mA cm−2 after long-term operation. For given temperatures, pressures, and humidity, which are considered fundamental factors in the hydrogen crossing phenomenon.The OHGR uses a stochastic gradient as an optimization engine that informs the most-recommended values for temperature, pressure, and humidity to generate the optimal amount of hydrogen. In addition, the OHGR model is interoperated to accept the same number of most significant features obtained from the principal component analysis. The OHGR was evaluated using empirical regression evaluation metrics, including the root mean square error, R2, Mean Square Error, and Mean Absolute Error. The optimization process was designed to include the hypothesis and hyperparameters to determine the most significant values for the OGHR and its outcomes. The introduced OHGR model is sufficiently efficient to predict the generated hydrogen with RMSE = 0.220 and R2 = 0.564, indicating an enhancement of the OHGR model compared to recently reported models. Since the sustainability of hydrogen generation is important for energy availability and reducing climate change, the OHGR model is justified using the SHapley Additive exPlanations (SHAP) method to ensure that the model is transparent, reliable, and trustworthy.

Enhancing Fermentation Process Monitoring through Data-Driven Modeling and Synthetic Time Series Generation.

Soft sensors based on deep learning regression models are promising approaches to predict real-time fermentation process quality measurements. However, experimental datasets are generally sparse and may contain outliers or corrupted data. This leads to insufficient model prediction performance. Therefore, datasets with a fully distributed solution space are required that enable effective exploration during model training. In this study, the robustness and predictive capability of the underlying model of a soft sensor was improved by generating synthetic datasets for training. The monitoring of intensified ethanol fermentation is used as a case study. Variational autoencoders were employed to create synthetic datasets, which were then combined with original datasets (experimental) to train neural network regression models. These models were tested on original versus augmented datasets to assess prediction improvements. Using the augmented datasets, the soft sensor predictive capability improved by 34%, and variability was reduced by 82%, based on R2 scores. The proposed method offers significant time and cost savings for dataset generation for the deep learning modeling of ethanol fermentation and can be easily adapted to other fermentation processes. This work contributes to the advancement of soft sensor technology, providing practical solutions for enhancing reliability and robustness in large-scale production.

Dual-stage deep learning for sangac optical fiber sensing multi-event detection and localization

This paper introduces a dual-stage deep learning structure for multi-event detection and localization (DDMDL) for a loop-based Sagnac interferometer optical fiber sensing system to overcome the difficulties of detecting and localizing multiple events. The DDMDL approach comprises of a combination of (i) a detection stage made of a multi-class classification model that quantifies the events, and (ii) a localization stage made of separate deep-learning regression models tasked with precisely localizing the events. By addressing the multi-event localization as a combined multi-class classification and multi-regression problem, our approach enables more accurate prediction of event locations. We evaluate our DDMDL model in a broad range of test scenarios across three signal-to-noise ratio (SNR) levels: low, medium, and high, all outside the model’s training dataset. The model had high detection (i.e., classification) accuracy in simulation, with rates of 90%–99% for single-event scenarios, 94%–98% for two-event scenarios, and 82%–99% for three-event scenarios at the three SNR levels. The precision of localization (i.e., regression) was evaluated by Mean Absolute Error (MAE). The results showed that single-event scenarios could be localized within a range of 35–50 m, two-event scenarios within a range of 82–106 m, and three-event scenarios within a range of 131–151 m. These findings were consistent across different SNR levels, ranging from low to high, over a 50 km sensing fiber length. Experimental validations confirmed the DDMDL model’s practical applicability, with detection accuracy of 80% in single-event scenarios and localization accuracy with MAEs ranging between 32 and 65 m. For two-event scenarios, the model achieved an 87% success rate, with MAEs ranging between 122 m and 204 m, emphasizing the model’s potential effectiveness in different applications.

Deep learning-based flocculation sensor for automatic control of flocculant dose in sludge dewatering processes during wastewater treatment

In sludge dewatering of most wastewater treatment plants (WWTPs), the dose of polymer flocculant is manually adjusted through direct visual inspection of the flocs without the aid of any instruments. Although there is a demand for the development of automatic control of flocculant dosing, this has been challenging owing to the lack of a reliable flocculation sensor. To address this issue, this study developed a novel image sensor for measuring the degree of flocculation (DF) based on deep learning. Two types of sludge samples were used in the laboratory-scale flocculation tests: excess sludge and mixtures of excess sludge and raw wastewater. To search for a deep learning regression model suitable for DF inference, ten models, including convolutional neural networks, vision transformers, and a multilayer perceptron MLP mixer, were comparatively analysed. The ConvNeXt and MLP mixer models showed the highest accuracies with coefficients of determination (R2) greater than 0.9. The region contributing to the DF inference in the flocculation images was visualised using a modulus-weighted, gradient-weighted regression activation map. A prototype of the flocculation sensor was constructed using an inexpensive EdgeAI device. This device comprises a single-board computer and an integrated graphical processing unit (GPU) and is equipped with a quantised ConvNeXt model. The maximum inference speed of the sensor was 12.8 frames per second (FPS). The flocculation control tests showed that the sensor could control the DF to the target value by adjusting the polymer flocculant dose. Therefore, the flocculation sensor can provide a data-driven approach to the management of the flocculation process, thereby facilitating the automation of WWTPs.

Symbolic regression as a feature engineering method for machine and deep learning regression tasks

In the realm of machine and deep learning (DL) regression tasks, the role of effective feature engineering (FE) is pivotal in enhancing model performance. Traditional approaches of FE often rely on domain expertise to manually design features for machine learning (ML) models. In the context of DL models, the FE is embedded in the neural network’s architecture, making it hard for interpretation. In this study, we propose to integrate symbolic regression (SR) as an FE process before a ML model to improve its performance. We show, through extensive experimentation on synthetic and 21 real-world datasets, that the incorporation of SR-derived features significantly enhances the predictive capabilities of both machine and DL regression models with 34%–86% root mean square error (RMSE) improvement in synthetic datasets and 4%–11.5% improvement in real-world datasets. In an additional realistic use case, we show the proposed method improves the ML performance in predicting superconducting critical temperatures based on Eliashberg theory by more than 20% in terms of RMSE. These results outline the potential of SR as an FE component in data-driven models, improving them in terms of performance and interpretability.

GraphEGFR: Multi-task and transfer learning based on molecular graph attention mechanism and fingerprints improving inhibitor bioactivity prediction for EGFR family proteins on data scarcity.

The proteins within the human epidermal growth factor receptor (EGFR) family, members of the tyrosine kinase receptor family, play a pivotal role in the molecular mechanisms driving the development of various tumors. Tyrosine kinase inhibitors, key compounds in targeted therapy, encounter challenges in cancer treatment due to emerging drug resistance mutations. Consequently, machine learning has undergone significant evolution to address the challenges of cancer drug discovery related to EGFR family proteins. However, the application of deep learning in this area is hindered by inherent difficulties associated with small-scale data, particularly the risk of overfitting. Moreover, the design of a model architecture that facilitates learning through multi-task and transfer learning, coupled with appropriate molecular representation, poses substantial challenges. In this study, we introduce GraphEGFR, a deep learning regression model designed to enhance molecular representation and model architecture for predicting the bioactivity of inhibitors against both wild-type and mutant EGFR family proteins. GraphEGFR integrates a graph attention mechanism for molecular graphs with deep and convolutional neural networks for molecular fingerprints. We observed that GraphEGFR models employing multi-task and transfer learning strategies generally achieve predictive performance comparable to existing competitive methods. The integration of molecular graphs and fingerprints adeptly captures relationships between atoms and enables both global and local pattern recognition. We further validated potential multi-targeted inhibitors for wild-type and mutant HER1 kinases, exploring key amino acid residues through molecular dynamics simulations to understand molecular interactions. This predictive model offers a robust strategy that could significantly contribute to overcoming the challenges of developing deep learning models for drug discovery with limited data and exploring new frontiers in multi-targeted kinase drug discovery for EGFR family proteins.

UAV remote sensing phenotyping of wheat collection for response to water stress and yield prediction using machine learning

Water stress is a significant challenge for global food production. Rainfall pattern is becoming unpredictable due to climate change that causes unprecedent water stress conditions in cereals production including wheat which is one of the important staple food crops. To sustain wheat production under water limiting conditions, there is an urgent need to develop drought-tolerant wheat varieties. For this, screening large numbers of wheat genotype for traits related to growth and yield under water stressed conditions is crucial. In this study, we deployed high-throughput phenotyping approaches, including uncrewed aerial vehicle (UAV)-based multispectral imaging, advanced machine and deep learning regression models. Two separate field experiments, irrigated and rainfed, were conducted comprising 553 wheat genotypes, and collected dataset for traits such as plant height, phenology, grain yield, and timeseries multispectral imaging. UAV-multispectral imagery derived plant height measurements showed a high correlation (R2=0.75) with manual measurements. Vegetation indices derived from multispectral data differentiated growth pattern of genotypes under rainfed and irrigated conditions and were used in yield prediction modeling. Wheat genotypes were effectively ranked, and their response differentiated for water stress tolerance based on yield index, stress susceptibility index, and yield loss%. Importantly, yield prediction in genotypes was computed using four machine learning regression algorithms i.e., linear regression, support vector machine, random forest, and deep learning H2O-3, where H2O-3 was the most accurate model with R2=0.80. Results show that multispectral-driven traits combined with machine learning models effectively phenotyped large wheat population and such approaches can be integrated in crop breeding program to develop varieties tolerant to water stress.

The rapid detection method of brown tide algae concentration based on unthresholded recurrence plots and improved broad learning system

In recent years, there have been frequent outbreaks of harmful algal blooms (HAB) in coastal areas, which have caused serious economic losses to the local community. Therefore, accurate and rapid prediction of microalgal concentrations is necessary for early warning and countermeasures before the occurrence of HAB. This paper presents a model for predicting microalgae concentration based on unthresholded recurrence plots (UTRPs) combined with an improved broad learning system (BLS). Spectral data acquisition of algae species at different concentrations using l-induced fluorescence spectroscopy. Then, the 1D spectral data are dimensionally lifted by UTRPs transformation, recurrence plots (RPs) transformation can fully extract the internal information of 1D sequence data, and at the same time, UTRPs avoids the influence of artificially selected thresholds on the feature transformation results of traditional (RPs). Finally, a lightweight flat network BLS was used for microalgae concentration regression prediction, at the same time the BLS regularization method was improved. When comparing the two most commonly used deep learning regression models and integrated learning models, UTRPs and elastic net (UTRPs-ENBLS) achieves convincing results.

External testing of an established glaucoma screening AI on a glaucoma clinic population from Finland

Purpose: To assess the external performance of an existing deep learning regression model for glaucoma detection (G‐RISK) in a glaucoma clinic population from Finland.Methods: Colour fundus images were retrospectively collected from the Tampere University Hospital (Tays), Finland. Eyes that had a visit in 2019 or 2020 were included based on the availability of ground truth labels. Ground truth labels for glaucoma were obtained through clinical evaluation of the optic nerve head (ONH) from colour fundus photos, clinical evaluation of retinal nerve fibre layer (RNFL) defects from fundus photos imaged with a cobalt filter, and clinical evaluation of the SITA Fast visual field (VF) test. Glaucoma diagnosis at Tays followed the ‘2 out of 3 rule’ recommended by the Finnish Evidence‐Based Guideline for Glaucoma, which requires at least 2 positive indications of glaucoma tests for referral [1]. Preprocessed colour fundus images were used as input to a previously trained deep learning model (G‐RISK) [2]. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), accompanied by 95% confidence intervals. Patient aggregated results were obtained by comparing the maximum of both prediction and ground truth label.Results: Over 3000 patients had at least one ground truth available for a visit in 2019 or 2020. Clinical evaluation of structural features (ONH, RNFL) triggered more glaucomatous symptoms than the SITA Fast VF test. When thresholding the AI prediction against a single ground truth type, the best AUC was obtained on the ONH label (0.90 [0.89–0.91]), followed by RNFL (0.84 [0.83–0.86]) and VF (0.83 [0.81–0.84]). When implementing the 2 out of 3 rule, AUC reached 0.88 [0.82–0.90].Conclusion: The model performed best when using the ONH label and incorporating the ‘2 out of 3 rule’, reaching an AUC of 0.88. These findings suggest that the G‐RISK model could be a valuable tool for screening in clinical practice.

Geospatial Big Data Approaches to Estimate Granular Level Poverty Distribution in East Java, Indonesia using Machine Learning and Deep Learning Regressions

One of the economic development the focus of the Indonesian government's efforts is for reducing poverty. In Indonesia, collecting poverty data uses the conventional method, the name is National Socio-Economic Survey (SUSENAS) which takes a large cost, time, and effort. To overcome these limitations, there is a need for additional data to provide more detailed poverty data. Recent studies show that the use of geospatial big data could identify poverty at a granular level, with a lower cost and faster update because of their unique and unbiased capacity to identify physical and socioeconomic phenomena. The integrated multi-source satellite imagery data such as the normalized difference vegetation index (NDVI) for detecting rural areas based on vegetation, built-up index (BUI) for identifying urban areas through building distribution, normalized difference water index (NDWI) for land cover detection, day time land surface temperature (LST) for identifying urban regions based on surface temperature, and pollutants such as carbon monoxide (CO), nitrogen dioxide (NO2), and sulfur dioxide (SO2) to evaluate economic activities based on pollution. Additionally, point of interest (POI) density and minimum POI distance are used to measure area accessibility. Therefore, the contribution of this research is to implement the utilization of geospatial big data to estimate the numbers of poverties at a granular level to the 666 sub-districts in East Java Province using machine learning and deep learning regression models. The evaluation results to estimate sub-district level poverty shows that the best model development using Support Vector Regression (SVR) in machine learning was the best root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) values of 0.365, 0.293, and 0.032 with R-squared of 0.59 and MLP in deep learning algorithm with 0.444, 0.345, and 0.039 values of RMSE, MAE, and MAPE with R2 0.52. In addition, the results of visual identification revealed that high estimates of lower poverty are typically found in urban areas with high accessibility, and these areas are not spatially deprived areas with limited accessibility.

Early identification of Alzheimer’s risk using anatomically interpretable estimation of brain age

AbstractBackgroundNeuroanatomical aging patterns can be revealed by examining the discrepancy between brain age (BA) and chronological age (CA) – also known as age gaps (AGs). Older than expected BAs, as estimated using deep learning of magnetic resonance images (MRIs), are observed in patients with mild cognitive impairment (MCI) and Alzheimer’s Disease (AD) (Tseng et al., 2022). Our 3‐dimensional convolutional neural network (3D‐CNN) enhances the accuracy of BA estimation and employs saliency mapping to explore structural patterns of aging between persons with and without CI.MethodsMRIs and neurocognitive measures for 6,561 participants (5,851 cognitively normal ‘CN’, 351 MCI, 359 AD) were collated across online repositories (Downey & Peakman, 2008; Petersen et al., 2010; Van Essen et al., 2012). We constructed a deep learning regression model using a 3D‐CNN whose inputs are pre‐processed MRIs (Fischl, 2012) and whose outputs are estimated BAs and saliency maps (Yin et al., 2023). Spearman rank correlations between neurocognitive measures and BA and CA were calculated, then compared using Fisher’s z‐test. The ability of AGs to predict conversion from MCI to AD was examined using logistic regression.ResultsIn participants with CI, older BA (but not CA) is significantly correlated with worse neurocognitive function across all measures except delayed verbal recall and logical memory (Table 1, Fig 1). AGs are positively associated with MCI participants’ probability of conversion to AD (β = 1.417, t 343 = 2.240, p = 0.025). Saliency maps reveal that, in CI participants, the 3D‐CNN relies more on features of the cerebral white matter, the brainstem, medial aspects of the temporal lobes, and the caudal portions of the anterior cingulate when estimating age according to cognitive status (Fig 2).ConclusionsCompared to CA, our BAs are significantly more strongly associated with early signs of Alzheimer’s disease, possibly reflecting disease‐related structural alterations. Saliency maps conveying the importance of each brain region for estimating BA reveal differences in patterns of neurological aging between persons with and without CI. Our BA estimation tool enhances the interpretability of such models, potentially improving the ability to stratify people according to risk of CI.

A Regression Framework for Energy Consumption in Smart Cities with Encoder-Decoder Recurrent Neural Networks

Currently, a smart city should ideally be environmentally friendly and sustainable, and energy management is one method to monitor sustainable use. This research project investigates the potential for a “smart city” to improve energy management by enabling the adoption of various types of intelligent technology to improve the energy sustainability of a city’s infrastructure and operational efficiency. In addition, the South Korean smart city region of Songdo serves as the inspiration for this case study. In the first module of the proposed framework, we place a strong emphasis on the data capabilities necessary to generate energy statistics for each of the numerous structures. In the second phase of the procedure, we employ the collected data to conduct a data analysis of the energy behavior within the microcities, from which we derive characteristics. In the third module, we construct baseline regressors to assess the proposed model’s varying degrees of efficacy. Finally, we present a method for building an energy prediction model using a deep learning regression model to solve the problem of 48-hour-ahead energy consumption forecasting. The recommended model is preferable to other models in terms of R2, MAE, and RMSE, according to the study’s findings.

Significant wave height prediction from X-band marine radar images using deep learning with 3D convolutions.

This research introduces a deep learning method for ocean wave height estimation utilizing a Convolutional Neural Network (CNN) based on the VGGNet. The model is trained on a dataset comprising buoy wave heights and radar images, both critical for marine engineering. The dataset features X-band radar images sourced from Sokcho, Republic of Korea, spanning from June 1, 2021, to August 13, 2021. This collection amounts to 72,180 three-dimensional images, gathered at intervals of approximately 1.43 seconds. The data collected was highly unbalanced in terms of wave heights, with images of lower wave heights being more common. To deal with data imbalances in the wave height datasets, we categorized the data into three groups based on wave heights and applied stratified random sampling at each level. This approach balances the data patches for each training iteration, reducing the risk of overfitting and promoting learning from diverse data. We also implemented a system to protect data in groups with fewer instances, ensuring fair representation across all categories. This study presents a deep learning regression model for predicting wave height values from radar images. The model extracts features from sequences of 64 radar images using three-dimensional convolutions for both temporal and spatial learning. Using three-dimensional convolutions, the model captures temporal features in radar image sequences and provides accurate wave height estimates with an RMSE of 0.3576 m. The study derived results using radar images under different wave height conditions for 74 days to ensure reliability.

Automatic Assessment of Depression and Anxiety through Encoding Pupil-wave from HCI in VR Scenes

At present, there have been many studies on the methods of using the deep learning regression model to assess depression level based on behavioral signals (facial expression, speech, and language); however, the research on the assessment method of anxiety level using deep learning is absent. In this article, pupil-wave, a physiological signal collected by Human Computer Interaction (HCI) that can directly represent the emotional state, is developed to assess the level of depression and anxiety for the first time. In order to distinguish between different depression and anxiety levels, we use the HCI method to induce the participants’ emotional experience through three virtual reality (VR) emotional scenes of joyful, sad, and calm, and construct two differential pupil-waves of joyful and sad with the calm pupil-wave as the baseline. Correspondingly, a dual-channel fusion depression and anxiety level assessment model is constructed using the improved multi-scale convolution module and our proposed width-channel attention module for one-dimensional signal processing. The test results show that the MAE/RMSE of the depression and anxiety level assessment method proposed in this article is 3.05/4.11 and 2.49/1.85, respectively, which has better assessment performance than other related research methods. This study provides an automatic assessment technique based on human computer interaction and virtual reality for mental health physical examination.

Multivariate multi-step LSTM model for flood runoff prediction: a case study on the Godavari River Basin in India

Abstract Flood is India's most prevalent natural calamity, devastatingly affecting human lives, infrastructure, and agriculture. Predicting floods can help to mitigate the potential damage and conduct timely evacuation drives. This research proposes a deep-learning regression model to forecast flood runoff. Various climatological, hydrological, land, and vegetation-related data have been collected from multiple sources for 18 years (2002–2019) to create a comprehensive dataset for the Godavari River at the Perur water station in India. The relevant attributes identified through feature selection are river water level, precipitation, temperature, surface pressure, evaporation, soil water content, daily runoff, and average river flow. The selected features were fed into various time series prediction models like AutoRegressive Integrated Moving Average (ARIMA), Prophet, Neural Prophet, and Long Short-Term Memory (LSTM). The LSTM model obtained the best results achieving a Root Mean Squared Error (RMSE) value of 0.05, Mean Absolute Error (MAE) value of 0.007, Willmott's Index (WI) of 0.83, Legates-McCabe's Index (LMI) of 0.58, and R2 of 0.67 for a 1-day prediction with a look-back window of 183 days. The model is also trained to predict the flood runoff value for a week ahead. The proposed model can serve as an essential component in flood warning systems.

Leaf water content determination of oilseed rape using near-infrared hyperspectral imaging with deep learning regression methods

Water content is crucial for plant growth. Determination of water content can help monitor plant growth status. In this study, spectral data in the range of 900–1700 nm acquired by near-infrared hyperspectral imaging and corrected by black-white calibration were used to detect the water content of fresh oilseed rape leaves. The oilseed leaves were analyzed without particular treatments. Conventional machine learning (support vector regression, partial least squares regression and least absolute shrinkage and selection operator) and deep learning regression models (Convolutional Neural Network and Long Short-Term Memory) were developed to predict oilseed rape leaf water content. The performance of CNN-LSTM-R was highly accurate. The coefficient of determination and root mean square error of the testing set (RMSEP) were 0.814 and 0.005, respectively. The characteristic wavelengths with strong correlation with water content prediction of the regression models were analyzed. The results showed that the deep learning-based regression models showed great potential for water content determination of oilseed rape leaves. Therefore, this study provides an important theoretical basis and practical application for the detection of fresh plant water content.

Evolutionary conditional GANs for supervised data augmentation: The case of assessing berry number per cluster in grapevine

In viticulture, training a deep learning model to assess the number of berries in a cluster image requires labeling by manually counting the berries on each grape cluster. This repeating, demanding task directly opposes the need for large amounts of data of deep learning methods. The objective of this work is the development of evolutionary conditional generative adversarial networks (GANs) for supervised data augmentation in the task of assessing berry number per cluster in grapevine. Ninety-seven grape cluster images were collected and labeled by manually counting the berries, making up the original dataset, and it was used for the training of a conditional GAN using evolutionary schemes involving a population of generators competing within a common and dynamic environment: the discriminator. After generative networks training, the best performing generator was used for supervised data augmentation: generation of labeled images conditioned to a berry number, making up the augmented dataset. Two models were trained with one dataset each, original and augmented, both having approximately the same number of samples, around 400. The original dataset was enriched with traditional image transforms, while the augmented dataset had incorporated images from the trained GAN generator. The two models were trained using the same convolutional regression network architecture, and then tested on an external image dataset, with more than 1300 images not used in any training, to compare the performance of GAN data augmentation. Results showed that the model augmented with GANs yielded lower error values, with a validation error of 43 berries, than those from the original model, with a validation error of 65 berries. In the test dataset, the augmented model obtained an error of 39 berries (R2 = 0.75), surpassing again the original model, that obtained an error of 57 berries (R2 = 0.65). These results evidence that evolutionary conditional GANs generate synthetic labeled images that lead to higher performance deep learning regression models for assessing berry number from cluster images.

MnEdgeNet for accurate decomposition of mixed oxidation states for Mn XAS and EELS L2,3 edges without reference and calibration

Accurate decomposition of the mixed Mn oxidation states is highly important for characterizing the electronic structures, charge transfer and redox centers for electronic, and electrocatalytic and energy storage materials that contain Mn. Electron energy loss spectroscopy (EELS) and soft X-ray absorption spectroscopy (XAS) measurements of the Mn L2,3 edges are widely used for this purpose. To date, although the measurements of the Mn L2,3 edges are straightforward given the sample is prepared properly, an accurate decomposition of the mix valence states of Mn remains non-trivial. For both EELS and XAS, 2+, 3+, and 4+ reference spectra need to be taken on the same instrument/beamline and preferably in the same experimental session because the instrumental resolution and the energy axis offset could vary from one session to another. To circumvent this hurdle, in this study, we adopted a deep learning approach and developed a calibration-free and reference-free method to decompose the oxidation state of Mn L2,3 edges for both EELS and XAS. A deep learning regression model is trained to accurately predict the composition of the mix valence state of Mn. To synthesize physics-informed and ground-truth labeled training datasets, we created a forward model that takes into account plural scattering, instrumentation broadening, noise, and energy axis offset. With that, we created a 1.2 million-spectrum database with 1-by-3 oxidation state composition ground truth vectors. The library includes a sufficient variety of data including both EELS and XAS spectra. By training on this large database, our convolutional neural network achieves 85% accuracy on the validation dataset. We tested the model and found it is robust against noise (down to PSNR of 10) and plural scattering (up to t/λ = 1). We further validated the model against spectral data that were not used in training. In particular, the model shows high accuracy and high sensitivity for the decomposition of Mn3O4, MnO, Mn2O3, and MnO2. The accurate decomposition of Mn3O4 experimental data shows the model is quantitatively correct and can be deployed for real experimental data. Our model will not only be a valuable tool to researchers and material scientists but also can assist experienced electron microscopists and synchrotron scientists in the automated analysis of Mn L edge data.