Machine Learning Model Selection Research Articles

Detection of basal cell carcinoma by machine learning-assisted ex vivo confocal laser scanning microscopy.

Ex vivo confocal laser scanning microscopy (EVCM) is an emerging imaging modality that enables near real-time histology of whole tissue samples. However, the adoption of EVCM into clinical routine is partly limited because the recognition of modality-specific diagnostic features requires specialized training. Therefore, we aimed to build a machine learning algorithm for the detection of basal cell carcinoma (BCC) in images acquired using EVCM and, in turn, facilitate the examiner's decision-making process. In this proof-of-concept study, histologically confirmed BCC fresh tissue samples were used to generate 50 EVCM images to train and assess a convolutional neural network architecture (MobileNet-V1) via tenfold cross-validation. Overall sensitivity and specificity of the model for detecting BCC and tumor-free regions on EVCM images compared to expert evaluation were 0.88 and 0.85, respectively. We constructed receiver operator characteristic and precision-recall curves from the aggregated tenfold cross-validation to assess the model's performance; the area under the curve was 0.94 and 0.87, respectively. Subsequently, the performance of one of the selected machine learning models was assessed with 19 new EVCM images of tumor-containing (n = 10) and 9 tumor-free (n = 9) skin tissue. A sensitivity of 0.83 and a specificity of 0.92 were achieved for the BCC group. The specificity for the tumor-free control group was 0.98. The deep learning model developed in our study holds the potential to assist the diagnostic decision-making process and diminish the training time of novices by depicting tumor-positive regions in EVCM images.

Application of Machine Learning in Construction Productivity at Activity Level: A Critical Review

There are two crucial resources (i.e., labor and equipment) of productivity in the construction industry. Productivity modeling of these resources would aid stakeholders in project management and improve construction scheduling and monitoring. Hence, this research aims to review machine learning (ML) applications in the process of construction productivity modeling (CPM) for construction labor productivity (CLP) and construction equipment productivity (CEP) from dataset acquisition to data analysis and evaluation, which includes their trends and applicability. An extensive analysis of 131 journals focused on the application of machine learning in construction productivity (ML-CP) from 1990 to 2024 via a mixed review methodology (bibliometric analysis and systematic review) was conducted. It can be concluded that despite the rise in automated dataset collection, the traditional method has its advantages. The review further found that the selection of ML models relies on each particular application, available data, and computational resources. Noticeably, artificial neural networks, convolutional neural networks, support vector machines, and even deep learning demonstrating have been adopted due to their effectiveness in different functionalities and processes in CPM. This study will supplement the insights gained in the review with a comprehensive understanding of how ML applications operate at each stage of CPM, enabling researchers to make future improvements.

Data augmentation using SMOTE technique: Application for prediction of burst pressure of hydrocarbons pipeline using supervised machine learning models

Accurate burst pressure prediction is critical for ensuring oil and gas pipeline safety, guiding maintenance decisions, and lowering costs and risks. Traditional methods have limitations, including high experimental costs, conservative empirical models, and computationally expensive numerical algorithms. Machine learning (ML) models have supplanted traditional methods in recent years. However, small and imbalanced datasets are the big challenge to build a ML model that can generate more accurate results. Moreover, the lack of generalization in ML models trained on a dataset of pipelines with specific material grids prevents them from producing superior results on other pipeline types. First, FEA was used to make a dataset. Then, a new way to improve machine learning (ML) model generalization for burst pressure prediction is suggested: combine publicly available datasets of different pipeline specifications. In this combined dataset, some pipelines have a higher number of data samples, and some have fewer, which causes a class imbalance issue. The Synthetic Minority Oversampling Technique (SMOTE) technique was applied to address the issue of class imbalance. The performance of various ML models, Extra Trees (ET), Extreme Gradient Boosting (XGBR), Random Forest (RF), Light Gradient Boosting Machine (LGBM), and Decision Tree (DT), was evaluated to validate the model's prediction and generalization on pipelines of various material grids. Results show that all the selected ML models produced high R-squared, i.e., >0.95, on balanced data compared to the imbalance dataset. These results show that SMOTE-based augmentation is a beneficial way to fix dataset imbalance and make ML models better at predicting burst pressure in oil and gas pipelines.

6999 Utilizing Machine Learning for One-Step Diagnosis Of Adrenal Diseases: Integrating Clinical, Serum Steroid, And Body Composition Data

Abstract Disclosure: S. Park: None. J. Noh: None. J. Kim: None. T. Kim: None. H. Seo: None. J. Choo: None. M. Choi: None. J. Kim: None. Background: With the increasing incidence of adrenal incidentalomas, the precise diagnosis of subtypes, such as mild autonomous cortisol secretion (MACS), adrenal Cushing’s syndrome (ACS), primary aldosteronism (PA), pheochromocytoma and paraganglioma (PPGL), or nonfunctioning adrenal adenomas (NFA) is becoming more critical. However, the process of making accurate diagnoses remains intricate and challenging. This study aims to explore a streamlined, one-step method for diagnosing adrenal nodules, employing a machine learning (ML) model that integrates clinical data (CL), serum steroid profiles (SSP), and computed tomography-based body composition data (BC). Methods: We utilized ML models for subtyping adrenal diseases within a cohort of 641 patients (comprising MACS=141, ACS=64, PA=265, PPGL=78, and NFA=93), incorporating 37 CL, 64 SSP, and 15 BC features. Patients were randomly split into training and test cohorts at a 4:1 ratio, with each disease proportionally represented. The ML-based pipeline involved ML model selection, outlier detection, feature selection, and re-sampling algorithm. An ML model was developed to classify each of the five distinct adrenal diseases individually. In addition, separate ML models were also developed and optimized for individual comparisons between NFA and each of the other four diseases. The performance was evaluated using balanced accuracy, sensitivity, specificity, area under the curve (AUC), and F1 score. Results: The OneVsRest Classifier exhibited the best performance in the multi-class classification of the five adrenal diseases, achieving a balanced accuracy (0.78), sensitivity (0.77), specificity (0.93), and an AUC of 0.89. In this process, 10 CL features and 39 SSP features were utilized, whereas BC did not yield significant features. Using only SSP, the ML model achieved a balanced accuracy of 0.71 and an AUC of 0.89, whereas the ML model using only CL demonstrated a relatively lower accuracy of 0.58 and an AUC of 0.82. In models comparing MACS, ACS, PA, and PPGL with NFAs, the accuracies were 0.85, 0.94, 0.78, and 0.86, respectively, with AUCs of 0.96, 0.99, 0.90, and 0.94, respectively. The ML model differentiating between NFA and the other functioning adrenal diseases exhibited an accuracy of 0.75 and an AUC of 0.79 Conclusion: We have developed a highly sensitive and specific one-step ML model for differentiating adrenal diseases using CL, SSP, and BC. This approach is anticipated to facilitate precise diagnoses of adrenal nodules, thereby enhancing patient care and improving clinical outcomes. Presentation: 6/2/2024

Landslide Recognition Based on Machine Learning Considering Terrain Feature Fusion

Landslides are one of the major disasters that exist worldwide, posing a serious threat to human life and property safety. Rapid and accurate detection and mapping of landslides are crucial for risk assessment and humanitarian assistance in affected areas. To achieve this goal, this study proposes a landslide recognition method based on machine learning (ML) and terrain feature fusion. Taking the Dawan River Basin in Detuo Township and Tianwan Yi Ethnic Township as the research area, firstly, landslide-related data were compiled, including a landslide inventory based on field surveys, satellite images, historical data, high-resolution remote sensing images, and terrain data. Then, different training datasets for landslide recognition are constructed, including full feature datasets that fusion terrain features and remote sensing features and datasets that only contain remote sensing features. At the same time, different ratios of landslide to non-landslide (or positive/negative, P/N) samples are set in the training data. Subsequently, five ML algorithms, including Extreme Gradient Boost (XGBoost), Adaptive Boost (AdaBoost), Light Gradient Boost (LightGBM), Random Forest (RF), and Convolutional Neural Network (CNN), were used to train each training dataset, and landslide recognition was performed on the validation area. Finally, accuracy (A), precision (P), recall (R), F1 score (F1), and intersection over union (IOU) were selected to evaluate the landslide recognition ability of different models. The research results indicate that selecting ML models suitable for the study area and the ratio of the P/N samples can improve the A, R, F1, and IOU of landslide identification results, resulting in more accurate and reasonable landslide identification results; Fusion terrain features can make the model recognize landslides more comprehensively and align better with the actual conditions. The best-performing model in the study is LightGBM. When the input data includes all features and the P/N sample ratio is optimal, the A, P, R, F1, and IOU of landslide recognition results for this model are 97.47%, 85.40%, 76.95%, 80.95%, and 71.28%, respectively. Compared to the landslide recognition results using only remote sensing features, this model shows improvements of 4.51%, 35.66%, 5.41%, 22.27%, and 29.16% in A, P, R, F1, and IOU, respectively. This study serves as a valuable reference for the precise and comprehensive identification of landslide areas.

Machine learning allows expert level classification of intraoperative motor evoked potentials during neurosurgical procedures

ObjectiveTo develop and evaluate machine learning (ML) approaches for muscle identification using intraoperative motor evoked potentials (MEPs), and to compare their performance to human experts. BackgroundThere is an unseized opportunity to apply ML analytic techniques to the world of intraoperative neuromonitoring (IOM). MEPs are the ideal candidates given the importance of their correct interpretation during a surgical operation to the brain or the spine. In this work, we develop and test a set of different ML models for muscle identification using intraoperative MEPs and compare their performance to human experts. In addition, we provide a review of the available literature on current ML applications to IOM data in neurosurgery. MethodsWe trained and tested five different ML classifiers on a MEP database developed from six different muscles in patients who underwent brain or spinal cord surgery. MEPs were obtained by both transcranial (TES) and direct cortical stimulation (DCS) protocols. The models were evaluated within a single patient and on previously unseen patients, considering signals from TES and DCS both independently and mixed. Ten expert neurophysiologists classified a set of 50 randomly selected MEPs, and their performance was compared to the best performing model. ResultsA total of 25.423 MEPs were included in the study. Random Forest proved to be the best performing model with 99 % accuracy in the single patient dataset task and a 78 %–94 % accuracy range on previously unseen patients. The model performance was maximized by representing MEPs as a set of features typically employed in signal processing compared to traditional neurophysiological parameters. The classification ability of the Random Forest model between six different muscles and across different MEP acquisition modalities (79 %) significantly exceeded that of human experts (mean 48 %). ConclusionsCarefully selected ML models proved to have reliable capacity of extracting meaningful information to classify intraoperative MEPs using a limited number of features, proving robustness across patients and signal acquisition modalities, outperforming human experts, and with the potential to act as decision support systems to the IOM team. Such encouraging results lay the path to further explore the underlying nature of clinically important signals, with the aim to continue to produce useful applications to make surgeries safer and more efficient.

Machine learning‐driven implementation of workflow optimization in cloud computing for IoT applications

AbstractThe optimization of workflow scheduling in Internet of Things (IoT) environments presents significant challenges due to the dynamic and heterogeneous nature of these systems. Traditional techniques must often adapt to fluctuating network conditions and varying data loads. To address these limitations, we propose a novel approach that leverages Automated Machine Learning (AutoML) integrated with cloud computing to optimize workflow scheduling for IoT applications. Our solution automates machine learning model selection, training, and tuning, significantly enhancing computational efficiency and adaptability. Through extensive experimentation, we demonstrate that our AutoML‐driven approach surpasses conventional algorithms across several key metrics, including accuracy, computational efficiency, adaptability to dynamic environments, and communication efficiency. Specifically, our method achieves a scheduling accuracy improvement of up to 25%, a reduced computational overhead by 30%, and a 40% enhancement in adaptability under dynamic conditions. Furthermore, the scalability of our solution is critical in cloud computing contexts, enabling efficient handling of large‐scale IoT deployments by leveraging cloud resources for distributed processing. This scalability ensures that our approach can effectively manage increasing data volumes and device heterogeneity inherent in modern IoT systems.

Seismic Response Prediction of Porcelain Transformer Bushing Using Hybrid Metaheuristic and Machine Learning Techniques: A Comparative Study

Although seismic response predictions are widely used for engineering structures, their applications in electrical equipment are rare. Overstressing at the bottom of the porcelain insulators during seismic events has made power transformer bushings in substations prone to failure. Thus, this paper proposed and compared six integrated machine learning (ML) models for seismic stress response predictions for porcelain transformer bushings using easily monitored acceleration responses. Metaheuristic algorithms such as particle swarm optimization were employed for architecture tuning. Prediction accuracies for stress response values and classifications were evaluated. Finally, shaking table tests and simulation analyses for a 1100 kV bushing were implemented to validate the accuracy of the six ML models. The results indicated that the proposed ML models can quickly forecast the maximum stress experienced by a porcelain bushing during earthquakes. Swarm intelligence evolutionary technologies could quickly and automatically aid in the retrofitting of architecture for the ML models. The K-nearest neighbor regression model had the best level of prediction accuracy among the six selected ML models for experimental and simulation validations. ML prediction models have clear benefits over frequently used seismic analytical techniques in terms of speed and accuracy for post-earthquake emergency relief in substations.

Implementation of artificial intelligence methods to the processes of automated metrics forecasting for software systems development projects

The subject matter of the article is the process of automated forecasting of project metrics for software development projects that are typically subject to evaluation. It also covers AI methods and models that can be used to generate basic roadmap templates and operational work lists, as well as alternative estimates depending on the context. The goal of the work is to study the foundations of creating a system for automated predicting of alternative evaluations of a software product. The following tasks were solved in the article: determining the stages of evaluation related to the assessment of alternatives in the life cycle of a software development project; investigating the problems of predicting and the main factors affecting the final indicators; exploring predicting methods that can be used to implement multivariate assessment of a software development project. The following methods are used: methods for evaluating and predicting labor costs in software development projects, machine and deep learning, and assessing their effectiveness for solving the prediction problem. The following results were obtained: the conceptual foundations for creating automated evaluation and prediction systems based on the analysis of the effectiveness of selected machine learning models were determined, the areas of application for artificial intelligence methods in the process of evaluating software development project indicators were identified, the performance indicators of various machine learning models were assessed based on certain model evaluation parameters that characterize prediction accuracy; a conceptual architecture of a project roadmap generation software tool based on the GPT language model was proposed. Conclusions: the use of machine and deep learning methods can improve the accuracy of predictions for key project indicators, provide the possibility of flexible generation of various alternative roadmap templates and operational work lists, making the planning and management process more efficient and transparent under conditions of high uncertainty of project requirements.

Enhancing Cybersecurity through GAN-Augmented and Hybrid Feature Selection Machine Learning Models: A Case Study on EVSE Data

Enhancing Cybersecurity through GAN-Augmented and Hybrid Feature Selection Machine Learning Models: A Case Study on EVSE Data

AI for AI-based intrusion detection as a service: Reinforcement learning to configure models, tasks, and capacities

Intrusion Detection Systems (IDS) increasingly leverage machine learning (ML) to enhance the detection of zero-day attacks. As operational complexities increase, enterprises are turning to Intrusion Detection as a Service (IDaS), requiring advanced solutions for efficient ML model selection and resource allocation. Existing research often focuses primarily on accuracy and computational efficiency, leaving a gap in solutions that can dynamically adapt. This study introduces a novel integrated solution, Auto-IDaS, which employs advanced Reinforcement Learning (RL) techniques for real-time, adaptive management of IDS. Auto-IDaS uses the Deep Q-Network (DQN) algorithm for dynamic ML model selection, automatically adjusting configurations of IDaS in response to fluctuating network traffic conditions. Simultaneously, it utilizes the Twin Delayed Deep Deterministic (TD3) algorithm for optimizing capacity allocation, aiming to minimize computational costs while maintaining service quality. This dual approach is innovative in its use of RL to address both selection and allocation challenges within IDaS frameworks. The effectiveness of TD3 is compared against Simulated Annealing (SA), a traditional optimization technique. The results demonstrate that utilizing DQN to dynamically select the model significantly improves the reward by 0.29% to 27.04%, effectively balancing detection performance (F1 score), detection time, and computation cost. Regarding capacity allocation, TD3 accelerates decision times approximately 5×106 times faster than SA while retaining decision quality within a 10% range comparable to SA’s performance.

Determining water and solute permeability of reverse osmosis membrane using a data-driven machine learning pipeline

Water (A) and solute permeability (B) are the key parameters characterizing the performance of reverse osmosis (RO) membranes. However, determining A and B requires multiple times of experiments that are time-consuming. Thus, developing a method to estimate A and B quickly has been encouraged. This study introduces a data-driven machine learning (ML) pipeline designed to predict A and B for RO membranes. The ML pipeline addresses all stages for estimating A and B, from data preprocessing to practical guidance on developed models. In particular, the membrane composition data, an overlooked essential component detailing A and B, was employed as features for the ML models. Sixteen ML models were tested, and categorical boost (CatBoost) and extremely randomized trees (ET) were selected as the best models for A and B, respectively. The selected ML models were retrained using the minimum yet effective features identified through feature importance analysis. Consequently, the retrained ML models exhibited high accuracy in predicting A and B (Radj,test2 = 0.925/MAEtest = 0.246 for A; Radj,test2 = 0.986/MAEtest = 0.069 for B), demonstrating the viability of a data-driven approach for parameter estimation. Moreover, this study highlights that RO membrane performance can be gauged effectively using membrane composition data and a limited set of operating conditions.

Machine Learning Approaches to Predict the Selectivity of Compounds against HDAC1 and HDAC6

The design of compounds selectively binding to specific isoforms of histone deacetylases (HDACs) is ongoing research to prevent adverse side effects. Two of the most studied isoforms are HDAC1 and HDAC6 which are important targets in various disease conditions. Here, various machine learning (ML) approaches were built and tested to predict the bioactivity and selectivity towards specific isoforms. The selectivities of compounds were precited using two different approaches: selectivity profiling and selectivity window approaches. The bioactivity models of HDAC1 and HDAC6 were used to determine the selectivities of compounds in the selectivity profiling approach. In the selectivity window approach, models were developed by directly training on the bioactivity differences of tested compounds against HDAC1 and HDAC6. In this study, firstly all available classification and regression ML algorithms in Python package were tested to find suitable algorithms. Five ML algorithms were selected based on their performances and algorithm differences. These models were compared to each other by using traditional evaluation metrics. Then, a consensus approach of the selected ML models was employed to compare the selectivity window and selectivity profiling approaches regarding their ability to distinguish HDAC1- and HDAC6-selective compounds from others. Here, the performances were also tested against an external set and it was seen that the selectivity window approach was performing slightly better in categorizing selective compounds than the selectivity profiling approach. The approaches presented in this study could be an important step to utilize for screening molecular libraries to discover selective inhibitors for targets.

Three-dimensional reconstruction of subsurface stratigraphy using machine learning with neighborhood aggregation

In engineering geology and geotechnical engineering, it is well recognized that subsurface soils/rocks are natural materials and exhibit variability in stratigraphy due to the complex geological formation processes they have undergone. Knowledge of subsurface soil stratigraphy is of great importance to geotechnical engineers. However, accurate and reliable interpretation of subsurface soil stratigraphy is challenging due to the limited number of site investigation boreholes available at the site and the highly heterogeneous properties of soil stratigraphy (e.g., interbedded or non-ordered layers). This paper proposes an improved data-driven machine learning framework boosted with the neighborhood aggregation technique for modelling three-dimensional (3D) subsurface soil stratigraphy in a more general and robust manner. Neighborhood aggregation, a technique often adopted in graph network learning, is integrated into this framework to regulate and improve the prediction results of classical machine learning models. The proposed framework is then cross-validated using 165 real site investigation boreholes for four selected machine learning models respectively. Cross-validation results suggest that the eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) models are more suitable for the task of soil stratigraphy prediction than Artificial Neural Network (ANN) and Support Vector Machine (SVM). Particularly, the XGBoost and RF are also amenable to neighborhood aggregation and can yield around 5% improvement in terms of average borehole prediction accuracy after introducing neighborhood aggregation. The improved machine learning framework allows for explicit 1D to 3D geological modelling, uncertainty quantification, and convenient visualization. The proposed framework facilitates digital transformation of geological and geotechnical site investigation.

Teaching classical machine learning as a graduate-level course in chemical engineering: An algorithmic approach

The demand for engineering graduates with technical skills in data science, machine learning (ML), and artificial intelligence (AI) is now growing. Chemical engineering (ChemE) departments around the world are currently addressing this skills gap by instituting AI or ML elective courses in their program. However, designing such a course is difficult since the issue of which ML models to teach and the depth of theory to be discussed remains unclear. In this paper, we present a graduate-level ML course particularly designed such that students will be able to apply ML for research in ChemE. To achieve this, the course intends to cover a wide selection of ML models with emphasis on their motivations, derivations, and training algorithms, followed by their applications to ChemE-related data sets. We argue that this algorithmic approach to teaching ML can help broaden the capabilities of students since they can judge for themselves which tool to use when, even for problems outside the process industries, or they can modify the methods to test novel ideas. We found that students remain engaged in the mathematical details as long as every topic is properly motivated and the gaps in the required statistical and computer science concepts are filled. Hence, this paper also presents a roadmap of ML topics, their motivations, and bridging topics that can be followed by instructors. Lastly, we report anonymized student feedback on this course which is being offered at the Department of Chemical Engineering, University of the Philippines, Diliman.

Subjectivity in Unsupervised Machine Learning Model Selection

Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants’ and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.

Strategic Selection of Machine Learning Models for Short-term Trading Optimization

Stock market prediction is a key aspect for a financial analyst and investor, as it facilitates better decisions and improved investment strategies. Towards this end, this project compares how machine learning (ML) models perform in predicting stock market prices with relatively small dataset sizes so as to get the best performing one. This study evaluates the efficiency and accuracy of artificial neural networks, support vector regression, LSTM and decision trees in capturing market trends and forecasting stock movements. The importance of our proposed paper lies in its potential to demystify the complexities of financial markets for students and new entrants in the field of finance. Participants can garner practical insights on how the stock market operates and the application of theoretical ML concepts to financial analytics by identifying the best-performing ML models with limited datasets, which is a common occurrence in real-life situations. This paper aims to analyse the commonly present machine learning models in stock market prediction and choose the most effective model depending upon their performance, which is achieved using limited information (short term data). This would provide an edge in taking small (less-risky) informed financial decisions without extensive study of the stock market.

Data-driven automatic classification model for construction accident cases using natural language processing with hyperparameter tuning

The construction industry, while vital to societal progress, is marred by a high incidence of accidents and injuries. Manual classification of accident cases is intensive and susceptible to human bias. This study addresses this challenge by developing an automated accident case classification system for the construction industry using Natural Language Processing and machine learning techniques. This study was conducted using the following steps: (1) Establishment of dataset (2) Korean Natural Language Processing (3) Selection of machine learning models (4) Model evaluation. The models exhibited competitive performance, demonstrating high accuracy, precision, and recall rates across all classification tasks. XGBoost outperformed NB, SVM, and KNN for accident type, facility type, and work type with accuracy of 0.80, 0.56, and 0.67, respectively. The results also provided insights into the factors influencing accident classification. This study contributes to construction safety by providing a data-driven foundation for safety decision-making, resource allocation, and benchmarking.

Prediction of flood sensitivity based on Logistic Regression, eXtreme Gradient Boosting, and Random Forest modeling methods.

Floods are one of the most destructive disasters that cause loss of life and property worldwide every year. In this study, the aim was to find the best-performing model in flood sensitivity assessment and analyze key characteristic factors, the spatial pattern of flood sensitivity was evaluated using three machine learning (ML) models: Logistic Regression (LR), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF). Suqian City in Jiangsu Province was selected as the study area, and a random sample dataset of historical flood points was constructed. Fifteen different meteorological, hydrological, and geographical spatial variables were considered in the flood sensitivity assessment, 12 variables were selected based on the multi-collinearity study. Among the results of comparing the selected ML models, the RF method had the highest AUC value, accuracy, and comprehensive evaluation effect, and is a reliable and effective flood risk assessment model. As the main output of this study, the flood sensitivity map is divided into five categories, ranging from very low to very high sensitivity. Using the RF model (i.e., the highest accuracy of the model), the high-risk area covers about 44% of the study area, mainly concentrated in the central, eastern, and southern parts of the old city area.

Evaluating machine learning models and identifying key factors influencing spatial maize yield predictions in data intensive farm management

Understanding the relationships between crop yields, soil properties, weather patterns and input applications is important for optimizing agricultural production. Data variation analysis using statistical and machine learning (ML) approaches can help identify and understand the practices that optimize yield. The objectives of this study were (i) to evaluate the predictive accuracy of selected ML models for estimating grain yields in on-farm maize (Zea mays L.) trials with different combinations of seeding and fertilizer rates in a commercial field, and (ii) to investigate the ability of ML models to assist in identifying yield-limiting factors in the same field. Multiple linear regression, multilayer perceptron, decision tree, and random forest (RF) ML models were trained and tested using crop management and soil from a data-intensive farm management (DIFM) trial and remotely sensed data. The dataset consisted of multiple subplot treatment observations of crop management, soil properties and normalized difference vegetation index (NDVI), linked to final grain yield for the 2019/2020 and 2020/2021 seasons. The RF had the best combination of high correlation (R2 = 0.69 and 0.80) and low error (MAPE = 5.4 and 8.4% and RMSE = 0.69 and 0.95 t ha−1) when compared to other models for both seasons. Feature importance analysis revealed that urea application was consistently the most critical variable and explained yield variations to the greatest extent, whereas soil phosphorus (P), plant population, and sodium in 2020, and soil P, soil pH, clay content, and plant population in 2021 emerged as the most influential factors for explaining yields. This study concluded that the RF model was the best for spatial yield predictions using DIFM trial datasets. There was also variability between seasons in yield limiting factors resulting from temporal variations in growing conditions. To effectively apply insights from yield prediction models, it is crucial that the variables incorporated into these models have a significant connection to yield and the findings can be translated into actionable management decisions. The DIFM trials combined with ML can play an important role in advancing the field of precision agriculture by providing valuable insights into the complex interactions between crops, soils, and management practices, and identifying new opportunities for improving crop yields and environmental sustainability.