Modern Machine Learning Methods Research Articles

Machine learning versus logistic regression for prognostic modelling in individuals with non-specific neck pain

PurposePrognostic models play an important clinical role in the clinical management of neck pain disorders. No study has compared the performance of modern machine learning (ML) techniques, against more traditional regression techniques, when developing prognostic models in individuals with neck pain.MethodsA total of 3001 participants suffering from neck pain were included into a clinical registry database. Three dichotomous outcomes of a clinically meaningful improvement in neck pain, arm pain, and disability at 3 months follow-up were used. There were 26 predictors included, five numeric and 21 categorical. Seven modelling techniques were used (logistic regression, least absolute shrinkage and selection operator [LASSO], gradient boosting [Xgboost], K nearest neighbours [KNN], support vector machine [SVM], random forest [RF], and artificial neural networks [ANN]). The primary measure of model performance was the area under the receiver operator curve (AUC) of the validation set.ResultsThe ML algorithm with the greatest AUC for predicting arm pain (AUC = 0.765), neck pain (AUC = 0.726), and disability (AUC = 0.703) was Xgboost. The improvement in classification AUC from stepwise logistic regression to the best performing machine learning algorithms was 0.081, 0.103, and 0.077 for predicting arm pain, neck pain, and disability, respectively.ConclusionThe improvement in prediction performance between ML and logistic regression methods in the present study, could be due to the potential greater nonlinearity between baseline predictors and clinical outcome. The benefit of machine learning in prognostic modelling may be dependent on factors like sample size, variable type, and disease investigated.

Sources of Risk of AI Systems.

Artificial intelligence can be used to realise new types of protective devices and assistance systems, so their importance for occupational safety and health is continuously increasing. However, established risk mitigation measures in software development are only partially suitable for applications in AI systems, which only create new sources of risk. Risk management for systems that for systems using AI must therefore be adapted to the new problems. This work objects to contribute hereto by identifying relevant sources of risk for AI systems. For this purpose, the differences between AI systems, especially those based on modern machine learning methods, and classical software were analysed, and the current research fields of trustworthy AI were evaluated. On this basis, a taxonomy could be created that provides an overview of various AI-specific sources of risk. These new sources of risk should be taken into account in the overall risk assessment of a system based on AI technologies, examined for their criticality and managed accordingly at an early stage to prevent a later system failure.

Age Classification in Forensic Medicine Using Machine Learning Techniques.

The aim of the study was to assess the capabilities of age determination (age group) at death using classification techniques by histomorphometric characteristics of osseous and cartilaginous tissue aging.Materials and MethodsThe study material was a database containing the findings of morphometric researches of osseous and cartilaginous tissue histologic specimens from 294 categorized male corpses aged 10–93 years. For data analysis and classification we used modern machine learning methods: k-NN, SVM, logistic regression, CatBoost, SGD, naive Bayes, random forest, nonlinear dimensionality reduction methods (t-SNE and uMAP), and recursive feature elimination for feature selection.ResultsThe used techniques (algorithms) provided effective representation of a complex data set (76 histomorphometric features), allowing to reveal the cluster structure inside the low dimensional feature space, thus fitting the classifier becomes even more reasonable. During feature selection, we estimated their importance for age group classification and studied the relationship between classification quality and the number of features inside the feature space. Data pre-processing made it possible to get rid of noise and keep most informative features, thereby accelerating a learning process and improving the classification quality. Data projection showed more well-defined cluster structure in the space of selected features. The accuracy of establishing certain groups was equal to 90%. It proves high efficiency of machine learning techniques used for forensic age diagnostics based on histomorphometric findings.

Raman Spectroscopy: A Personalized Decision-Making Tool on Clinicians' Hands for In Situ Cancer Diagnosis and Surgery Guidance.

Simple SummaryCancer still constitutes one of the main global health challenges. Novel approaches towards understanding the molecular composition of the disease can be employed as adjuvant tools to current oncological applications. Raman spectroscopy has been contemplated and pursued to serve as a noninvasive, real time, in vivo tool which may uncover the molecular basis of cancer and simultaneously offer high specificity, sensitivity, and multiplexing capacity, as well as high spatial and temporal resolution. In this review, the potential impact of Spontaneous Raman spectroscopy in clinical applications related to cancer diagnosis and surgical removal is analyzed. Moreover, the coupling of Raman systems with modern instrumentation and machine learning methods has been explored as a prominent enhancement factor towards a personalized approach promoting objectivity and accuracy in surgical oncology.Accurate in situ diagnosis and optimal surgical removal of a malignancy constitute key elements in reducing cancer-related morbidity and mortality. In surgical oncology, the accurate discrimination between healthy and cancerous tissues is critical for the postoperative care of the patient. Conventional imaging techniques have attempted to serve as adjuvant tools for in situ biopsy and surgery guidance. However, no single imaging modality has been proven sufficient in terms of specificity, sensitivity, multiplexing capacity, spatial and temporal resolution. Moreover, most techniques are unable to provide information regarding the molecular tissue composition. In this review, we highlight the potential of Raman spectroscopy as a spectroscopic technique with high detection sensitivity and spatial resolution for distinguishing healthy from malignant margins in microscopic scale and in real time. A Raman spectrum constitutes an intrinsic “molecular finger-print” of the tissue and any biochemical alteration related to inflammatory or cancerous tissue state is reflected on its Raman spectral fingerprint. Nowadays, advanced Raman systems coupled with modern instrumentation devices and machine learning methods are entering the clinical arena as adjunct tools towards personalized and optimized efficacy in surgical oncology.

Hybrid and automated machine learning approaches for oil fields development: The case study of Volve field, North Sea

The field development workflow contains numerous tasks involving decision-making processes. The modern machine learning methods, including automatic machine learning (AutoML), reduce the geophysics or machine learning experts’ time required to solve routine tasks. In the paper, we focus on the automated solution of the location of the wells optimization problem, namely, improving the quality of oil production estimation and estimating reservoir characteristics for appropriate wells placement and parametrization, using the same AutoML approach. Ideas of making several parallel or consequent tasks automatically within one framework are arising as Composite AI. We implemented and investigated the quality of forecasting models for oil production estimation: physics equation-based, pure data-driven, and hybrid. CRMIP (Capacitance–Resistance Model Injector–Producer) model is chosen as a physics-related approach. We automated the seismic analysis using evolutionary identification of convolutional neural network structure for reservoir detection to help investigate reservoir characteristics for wells location choice. The Volve oil field dataset was used as a case study to conduct the experiments. The implemented approaches can analyze different oil fields and even be adapted to similar physics-related problems.

The Role of Remote Sensing Data and Methods in a Modern Approach to Fertilization in Precision Agriculture

The precision fertilization system is the basis for upgrading conventional intensive agricultural production, while achieving both high and quality yields and minimizing the negative impacts on the environment. This research aims to present the application of both conventional and modern prediction methods in precision fertilization by integrating agronomic components with the spatial component of interpolation and machine learning. While conventional methods were a cornerstone of soil prediction in the past decades, new challenges to process larger and more complex data have reduced their viability in the present. Their disadvantages of lower prediction accuracy, lack of robustness regarding the properties of input soil sample values and requirements for extensive cost- and time-expensive soil sampling were addressed. Specific conventional (ordinary kriging, inverse distance weighted) and modern machine learning methods (random forest, support vector machine, artificial neural networks, decision trees) were evaluated according to their popularity in relevant studies indexed in the Web of Science Core Collection over the past decade. As a shift towards increased prediction accuracy and computational efficiency, an overview of state-of-the-art remote sensing methods for improving precise fertilization was completed, with the accent on open-data and global satellite missions. State-of-the-art remote sensing techniques allowed hybrid interpolation to predict the sampled data supported by remote sensing data such as high-resolution multispectral, thermal and radar satellite or unmanned aerial vehicle (UAV)-based imagery in the analyzed studies. The representative overview of conventional and modern approaches to precision fertilization was performed based on 121 samples with phosphorous pentoxide (P2O5) and potassium oxide (K2O) in a common agricultural parcel in Croatia. It visually and quantitatively confirmed the superior prediction accuracy and retained local heterogeneity of the modern approach. The research concludes that remote sensing data and methods have a significant role in improving fertilization in precision agriculture today and will be increasingly important in the future.

Benchmarking machine learning methods for modeling physical properties of ionic liquids

The great importance of the ability to quantitatively predict the properties of ionic liquids (ILs) using quantitative structure–property relationships (QSPR) models necessitates the understanding of which modern machine learning (ML) methods in combination with which types of molecular representations are preferable to use for this purpose. To address this problem, a large-scale benchmarking study of QSPR models built by combining three traditional ML methods and neural networks with seven different architectures with five types of molecular representations (in the form of either numerical molecular descriptors or SMILES text strings) to predict six important physical properties of ILs (density, electrical conductance, melting point, refractive index, surface tension, and viscosity) was carried out. The datasets include from 407 to 1204 diverse ILs composed of various organic and inorganic ions. QSPR models for predicting the properties of ILs at eight different temperatures were built using multi-task learning. The best combinations of ML methods and molecular representations were identified for each of the properties. A unified ranking system was introduced to rank and prioritize different ML methods and molecular representations. It was shown in this study that on average: (i) nonlinear ML methods perform much better than linear ones, (ii) neural networks perform better than traditional ML methods, (iii) Transformers, which are actively used in natural language processing (NLP), perform better than other types of neural networks due to the advanced ability to analyze chemical structures of ILs encoded into SMILES text strings. A special “component-wise” cross-validation scheme was applied to assess how much the predictive performance deteriorates for the ILs composed of cations and anions that are not present in the dataset.

Machine learning for predicting 5-year mortality risks: data from the ESSE-RF study in Primorsky Krai

Aim. To develop and perform comparative assessment of the accuracy of models for predicting 5-year mortality risks according to the Epidemiology of Cardiovascular Diseases and their Risk Factors in Regions of Russian Federation (ESSE-RF) study in Primorsky Krai.Material and methods. The study included 2131 people (1257 women and 874 men) aged 23-67 years with a median of 47 years (95% confidence interval [46; 48]). The study protocol included measurement of blood pressure (BP), heart rate (HR), waist circumference, hip circumference, and waist-to-hip ratio (WHR). The following blood biochemical parameters: total cholesterol (TC), low and high density lipoprotein cholesterol, triglycerides, apolipoproteins AI and B, lipoprotein(a), N-terminal pro-brain natriuretic peptide (NT-proNBP), D-dimer, fibrinogen, C-reactive protein (CRP), glucose, creatinine, uric acid. The study endpoint was 5-year all-cause death (2013-2018). The group of deceased patients during this period consisted of 42 (2%) people, while those continued the study — 2089 (98%). The χ2, Fisher and MannWhitney tests, univariate logistic regression (LR) were used for data processing and analysis. To build predictive models, we used following machine learning (ML) methods: multivariate LR, Weibull regression, and stochastic gradient boosting.Results. The prognostic models developed on the ML basis, using parameters of age, sex, smoking, systolic blood pressure (SBP) and TC level in their structure, had higher quality metrics than Systematic COronary Risk Evaluation (SCORE) system. The inclusion of CRP, glucose, NT-proNBP, and heart rate into the predictors increased the accuracy of all models with the maximum rise in quality metrics in the multivariate LR model. Predictive potential of other factors (WHR, lipid profile, fibrinogen, D-dimer, etc.) was low and did not improve the prediction quality. An analysis of the influence degree of individual predictors on the mortality rate indicated the prevailing contribution of five factors as follows: age, levels of TC, NT-proNBP, CRP, and glucose. A less noticeable effect was associated with the level of HR, SBP and smoking, while the contribution of sex was minimal.Conclusion. The use of modern ML methods increases the accuracy of predictive models and provides a higher efficiency of risk stratification, especially among individuals with a low and moderate death risk from cardiovascular diseases.

Forecasting patient arrivals at emergency department using calendar and meteorological information.

Overcrowding in emergency departments (EDs) is a serious problem in many countries. Accurate ED patient arrival forecasts can serve as a management baseline to better allocate ED personnel and medical resources. We combined calendar and meteorological information and used ten modern machine learning methods to forecast patient arrivals. For daily patient arrival forecasting, two feature selection methods are proposed. One uses kernel principal component analysis(KPCA) to reduce the dimensionality of all of the features, and the other is to use the maximal information coefficient(MIC) method to select the features related to the daily data first and then perform KPCA dimensionality reduction. The current study focuses on a public hospital ED in Hefei, China. We used the data November 1, 2019 to August 31, 2020 for model training; and patient arrival data September 1, 2020 to November 31, 2020 for model validation. The results show that for hourly patient arrival forecasting, each machine learning model has better forecasting results than the traditional autoRegressive integrated moving average (ARIMA) model, especially long short-term memory (LSTM) model. For daily patient arrival forecasting, the feature selection method based on MIC-KPCA has a better forecasting effect, and the simpler models are better than the ensemble models. The method we proposed could be used for better planning of ED personnel resources.

Review on Epileptic Seizure Prediction: Machine Learning and Deep Learning Approaches.

Epileptic seizures occur due to brain abnormalities that can indirectly affect patient's health. It occurs abruptly without any symptoms and thus increases the mortality rate of humans. Almost 1% of world's population suffers from epileptic seizures. Prediction of seizures before the beginning of onset is beneficial for preventing seizures by medication. Nowadays, modern computational tools, machine learning, and deep learning methods have been used to predict seizures using EEG. However, EEG signals may get corrupted with background noise, and artifacts such as eye blinks and physical movements of muscles may lead to “pops” in the signal, resulting in electrical interference, which is cumbersome to detect through visual inspection for longer duration recordings. These limitations in automatic detection of interictal spikes and epileptic seizures are preferred, which is an essential tool for examining and scrutinizing the EEG recording more precisely. These restrictions bring our attention to present a review of automated schemes that will help neurologists categorize epileptic and nonepileptic signals. While preparing this review paper, it is observed that feature selection and classification are the main challenges in epilepsy prediction algorithms. This paper presents various techniques depending on various features and classifiers over the last few years. The methods presented will give a detailed understanding and ideas about seizure prediction and future research directions.

On Residual-based Diagnosis of Physical Systems

In this article we describe a novel diagnosis methodology for physical systems such as industrial production systems. The article consists of two parts: Part one analyzes the differences between using sensor values and using residual values for fault diagnosis. Residual values denote the health of a component by comparing sensor values to a predefined model of normal behaviour. We further analyse how faults propagate through components of a physical system and argue for the use of residual values for diagnosing physical systems. In part two we extend the theory of established consistency-based diagnosis algorithms to use residual values. We also illustrate how users of the presented diagnosis methodology are free to substitute the residual generating equations and the diagnosis algorithm to suit their specific needs.For diagnosis, we present the algorithm HySD, based on Satisfiability Modulo Linear Arithmetic. We present an implementation of HySD using threshold values and a symbolic diagnosis approach. However, the approach is also suitable to integrate modern machine learning methods for anomaly detection and combine them with a multitude of diagnosis approaches.Through experiments on the process-industry benchmark Tennessee Eastman Process and another benchmark consisting of multiple tank systems we show the feasibility of our approach. Overall we show how our novel diagnosis approach offers a practical methodology that allows industry to advance from current state of the art anomaly detection to automated fault diagnosis.

Modern machine learning for tackling inverse problems in chemistry: molecular design to realization.

The discovery of new molecules and materials helps expand the horizons of novel and innovative real-life applications. In pursuit of finding molecules with desired properties, chemists have traditionally relied on experimentation and recently on combinatorial methods to generate new substances often complimented by computational methods. The sheer size of the chemical space makes it infeasible to search through all possible molecules exhaustively. This calls for fast and efficient methods to navigate the chemical space to find substances with desired properties. This class of problems is referred to as inverse design problems. There are a variety of inverse problems in chemistry encompassing various subfields like drug discovery, retrosynthesis, structure identification, etc. Recent developments in modern machine learning (ML) methods have shown great promise in tackling problems of this kind. This has helped in making major strides in all key phases of molecule discovery ranging from in silico candidate generation to their synthesis with a focus on small organic molecules. Optimization techniques like Bayesian optimization, reinforcement learning, attention-based transformers, deep generative models like variational autoencoders and generative adversarial networks form a robust arsenal of methods. This highlight summarizes the development of deep learning to tackle a wide variety of inverse design problems in chemistry towards the quest for synthesizing small organic compounds with a purpose.

VALIDATION OF LAND DEGRADATION CARDS ON THE BASIS OF GEOSPATIAL DATA

Today there is a lot of satellite data and products based on it in the public domain. By integrating them with heterogeneous socio-economic information and soil maps, model biophysical data using modern machine learning methods and modern approaches to geospatial data processing, it becomes possible to create maps of land degradation. Considering that classification maps, productivity maps and deforestation maps are the main intellectual components to create a degradation map, it is these products that affect the overall reliability of the results. For their validation, the necessary quality metrics are determined in the work, and the corresponding calculations are made. To evaluate the land cover map, independent test data were used to build a confusion matrix, and the obtained areas of the main crops were compared with statistical data. Agricultural land productivity was estimated using time series land cover classification maps and Crop Growth Modeling System (CGMS) biophysical plant development, as well as biophysical plant growth parameters using satellite data and biophysical plant development models. The LAI Map Accuracy Assessment (based on CGMS) is based on the comparison of Leaf Area Index (LAI) values modeled using the CGMS software framework with LAI ground measurement data collected through ground surveys. Numerous experiments were carried out to assess the quality of models and the results of deforestation maps on an independent test sample, which was not used at the neural network training stage. Degradation maps for several years were also analyzed and their validation was carried out with respect to productivity, in particular for the region that has undergone significant changes for the territory of Ukraine.

Forecasting Milk Delivery to Dairy- How Modern Statistical and Machine Learning Methods Can Contribute

Forecasting Milk Delivery to Dairy- How Modern Statistical and Machine Learning Methods Can Contribute

Machine learning in house price analysis: regression models versus neural networks

The problem addressed in this paper is automatic house price determination using multiple regression models and machine learning. In the practice of real estate appraisal, discussions about automated valuation (AVM) are increasingly common. Resistance to modern machine learning methods stems from a low level of knowledge about these methods and, as a result, an unawareness to what extent and where they can support the classical process of estimating property value. The problem is much broader, because it affects many aspects of the use of machine learning (including neural networks) in the broader real estate market. The process of real estate management at each stage generates huge information resources, which are used at different levels and to different extents by entities operating in the real estate market. One of such professional groups are real estate appraisers, for whom intelligent systems for monitoring the market and providing necessary data are becoming increasingly common and sought after. In this context, a comparative study of two models: multiple regression and neural networks has been carried out. Both methods were used to determine house prices, based on the same set of input data, and in the next step the effects of using these models for an additional group of objects with known characteristics and transaction prices were compared. The multivariate regression model obtained in the study was of medium quality, but sufficient for the purpose of comparative analysis. In the case of neural networks, the highest quality model was not obtained for the study sample, despite normalizing the variables to the required interval (0;1). In both models, the prices for the control sample were overestimated in most cases. However, this does not deny the relevance of further research and attempts to teach neural networks using larger data sets, especially in the case of properties other than typical residential units or land for residential development. Machine learning methods can be extremely useful especially in the processes of general valuation. In these processes, large sets of properties are estimated in a short period of time and with the same methods.

Explaining the Effects of Clouds on Remote Sensing Scene Classification

Most of Earth is covered by haze or clouds, impeding the constant monitoring of our planet. Preceding works have documented the detrimental effects of cloud coverage on remote sensing applications and proposed ways to approach this issue. However, up-to-now, little effort has been spent on understanding how exactly atmospheric disturbances impede the application of modern machine learning methods to Earth observation data. Specifically, we consider the effects of haze and cloud coverage on a scene classification task. We provide a thorough investigation of how classifiers trained on cloud-free data fail once they encounter noisy imagery – a common scenario encountered when deploying pretrained models for remote sensing to real use cases. We show how and why remote sensing scene classification suffers from cloud coverage. Based on a multi-stage analysis, including explainability approaches applied to the predictions, we work out four different types of effects that clouds have on the scene prediction. The contribution of our work is to deepen the understanding of the effects of clouds on common remote sensing applications, and consequently guide the development of more robust methods.

ОБНАРУЖЕНИЕ ИНЦИДЕНТОВ ИНФОРМАЦИОННОЙ БЕЗОПАСНОСТИ НА ОСНОВЕ ТЕХНОЛОГИИ НЕЙРОННЫХ СЕТЕЙ

Objective: develop model, algorithmic and software for detecting in real time attempts to disrupt the correct functioning of critical information infrastructure systems with neural network technologies. Methods analysis of modern machine learning methods and neural network technologies, synthesis and modeling of correct behavior of programs, algorithmization of learning processes and application of neural networks, experimental studies of developed algorithms and programs on the stand. Study results: The characteristics of machine learning methods and neural network technologies used to detect software and technical impacts and information security incidents are given. The method for solving this problem based on neural networks with LSTM and FFN architectures has been developed. The description of the algorithm and fragments of the software implementation of the method in the programming languages Python3 and Go using Tensorflow and Keras libraries is given. An important advantage of the proposed approach is the possibility of adapting the neural network in the event of a change in the mode and conditions of operation of the system. The results obtained during the experiments indicate the possibility and expediency of using this approach to detect software and technical impacts on critical information infrastructure systems on a time scale close to real with a high level of reliability. Scientific novelty: consists in the application of deep learning technology based on a long-term short-term neural network LSTM, which has the ability to adapt to changing modes and conditions, to solve the problem of detecting signs of a violation of the correct functioning of nodes of information and telecommunications systems in real time.

Indirect supervision applied to COVID-19 and pneumonia classification

The novel coronavirus 19 (COVID-19) continues to have a devastating effect around the globe, leading many scientists and clinicians to actively seek to develop new techniques to assist with the tackling of this disease. Modern machine learning methods have shown promise in their adoption to assist the healthcare industry through their data and analytics-driven decision making, inspiring researchers to develop new angles to fight the virus. In this paper, we aim to develop a CNN-based method for the detection of COVID-19 by utilizing patients' chest X-ray images. Developing upon the inclusion of convolutional units, the proposed method makes use of indirect supervision based on Grad-CAM. This technique is used in the training process where Grad-CAM's attention heatmaps support the network's predictions. Despite recent progress, scarcity of data has thus far limited the development of a robust solution. We extend upon existing work by combining publicly available data across 5 different sources and carefully annotate the comprising images across three categories: normal, pneumonia, and COVID-19. To achieve a high classification accuracy, we propose a training pipeline based on indirect supervision of traditional classification networks, where the guidance is directed by an external algorithm. With this method, we observed that the widely used, standard networks can achieve an accuracy comparable to tailor-made models, specifically for COVID-19, with one network in particular, VGG-16, outperforming the best of the tailor-made models.

Artificial intelligence: machine learning for chemical sciences.

Research in molecular sciences witnessed the rise and fall of Artificial Intelligence (AI)/ Machine Learning (ML) methods, especially artificial neural networks, few decades ago. However, we see a major resurgence in the use of modern ML methods in scientific research during the last few years. These methods have had phenomenal success in the areas of computer vision, speech recognition, natural language processing (NLP), etc. This has inspired chemists and biologists to apply these algorithms to problems in natural sciences. Availability of high performance Graphics Processing Unit (GPU) accelerators, large datasets, new algorithms, and libraries has enabled this surge. ML algorithms have successfully been applied to various domains in molecular sciences by providing much faster and sometimes more accurate solutions compared to traditional methods like Quantum Mechanical (QM) calculations, Density Functional Theory (DFT) or Molecular Mechanics (MM) based methods, etc. Some of the areas where the potential of ML methods are shown to be effective are in drug design, prediction of high–level quantum mechanical energies, molecular design, molecular dynamics materials, and retrosynthesis of organic compounds, etc. This article intends to conceptually introduce various modern ML methods and their relevance and applications in computational natural sciences.Graphical abstract Synopsis Recent surge in the application of machine learning (ML) methods in fundamental sciences has led to a perspective that these methods may become important tools in chemical science. This perspective provides an overview of the modern ML methods and their successful applications in chemistry during the last few years.

Spatially Coherent Postprocessing of Cloud Cover Ensemble Forecasts

AbstractStatistical postprocessing is commonly applied to reduce location and dispersion errors of probabilistic forecasts provided by numerical weather prediction (NWP) models. If postprocessed forecast scenarios are required, the combination of ensemble model output statistics (EMOS) for univariate postprocessing with ensemble copula coupling (ECC) or the Schaake shuffle (ScS) to retain the dependence structure of the raw ensemble is a state-of-the-art approach. However, modern machine learning methods may lead to both a better univariate skill and more realistic forecast scenarios. In this study, we postprocess multimodel ensemble forecasts of cloud cover over Switzerland provided by COSMO-E and ECMWF-IFS using (i) EMOS + ECC, (ii) EMOS + ScS, (iii) dense neural networks (dense NN) + ECC, (iv) dense NN + ScS, and (v) conditional generative adversarial networks (cGAN). The different methods are verified using EUMETSAT satellite data. Dense NN shows the best univariate skill, but cGAN performed only slightly worse. Furthermore, cGAN generates realistic forecast scenario maps, while not relying on a dependence template like ECC or ScS, which is particularly favorable in the case of complex topography.