Feature Selection Method Research Articles

Comparison of different machine learning methods in river streamflow estimation using isovel contours and hydraulic variables

ABSTRACT This study evaluates several machine learning (ML) models for estimating streamflow in the Osage River, Missouri, USA, and the Severn River, UK, using hydraulic input variables. These input variables are the flow section area (A), wetted perimeter (Pw ), water surface width (W), and velocity parameter (U) derived from isovel contours. Before using these models, the influential variables are selected using the sequential feature selection (SFS) method to reduce model complexity while maintaining performance. The two most important input variables identified were A and U, reducing the number of inputs from four to two. The results showed that the river’s flow conditions affect the accuracy of ML models used for estimating streamflow. Linear models such as MLR and ANFIS perform better in steady flow conditions (Osage River). In contrast, decision tree-based and non-linear models such as SVR with a radial basis kernel function (SVR_RBF) are better for unsteady flow conditions (Severn River). The finding suggests simpler models outperform complex deep-learning approaches (such as LSTM) for estimating streamflow by hydraulic variables. By selecting appropriate and efficient ML models, hydrometer stations can significantly improve the accuracy of streamflow estimation, leading to more informed decision-making in water management.

An improved conditional relevance and weighted redundancy feature selection method for gene expression data

An improved conditional relevance and weighted redundancy feature selection method for gene expression data

Development of a spectral repository for the identification of western Himalayan medicinal plants using machine learning techniques

The identification of medicinal plant species is a crucial task for assessing the status of our bioresources. Conventional methods primarily rely on taxonomy and laboratory-based instruments, which are time-consuming and require the requisite expertise. Thus, there is an escalating demand for efficient techniques that can quickly identify these precious species. The advent of Hyperspectral Remote Sensing (HRS) with artificial intelligence has significantly increased the scope of HRS techniques by offering rapid and precise plant identification. This study utilised non-imaging HRS handheld sensors to build a spectral repository for 10 important medicinal plant species from diverse locations across Indian Himalayan states, representing varying altitudinal and ecological conditions. The spectral repository encompasses 1237 distinct spectral signatures obtained from the leaves and canopies of the targeted plant species. Subsequently, an identification model has been developed using Random Forest (RF) with several feature selection methods, and it has been revealed that the RF model, coupled with wrapper-based feature selection, is an effective combination for classifying the targeted plant species. The calibration and test datasets accounted for accuracies of 87.87% and 91.39%, respectively, with corresponding kappa coefficients of 0.85 and 0.89. Furthermore, the developed RF model was applied to ‘PRISMA’ satellite data to identify Saussurea costus crops in farmers' croplands, achieving a classification accuracy of 81.31% and a kappa coefficient of 0.76. Therefore, the study highlights the potential of integrating RF, in-situ HRS, and satellite HRS for the non-destructive, precise, and accurate identification of medicinal plants that can significantly contribute to biodiversity conservation and sustainable resource management.

Unveiling Lung Diseases in CT Scan Images With a Hybrid Bio‐Inspired Mutated Spider‐Monkey and Crow Search Algorithm

ABSTRACTBio‐inspired computer‐aided diagnosis (CAD) has garnered significant attention in recent years due to the inherent advantages of bio‐inspired evolutionary algorithms (EAs) in handling small datasets with elevated precision and reduced computational complexity. Traditional CAD models face limitations as they can only be developed post‐outbreak, relying on datasets that become available during such events such as the COVID‐19 pandemic. The scarcity of data for emerging diseases poses a substantial challenge to achieving elevated precision with conventional deep‐learning algorithms. Furthermore, even when datasets are available, employing deep learning for class‐based classification is arduous, necessitating model retraining, in this paper, we propose a novel hybrid algorithm that leverages the strengths of the crow search algorithm (CSA) and the spider monkey optimization (SMO) algorithm to create an optimised spider monkey crow search (OSM‐CS) algorithm. We developed a CAD tool that maps each input CT image to a high‐dimensional vector by extracting four categories of features: high contrast, polynomial decomposition, textural, and pixel statistics. The proposed OSM‐CS algorithm is employed as a feature selection method. Our experimental results demonstrate the effectiveness of the OSM‐CS algorithm, achieving an impressive accuracy of 98.2% when coupled with an AdaBoost classifier for multi‐class classification and 99.93% for binary classification. This performance surpasses that of state‐of‐the‐art (SOTA) deep learning models and recently published hybrid algorithms, underscoring the potential of the OSM‐CS algorithm as a powerful tool in the realm of CAD.

Enhanced Intrusion Detection in Software-Defined Networking using Advanced Feature Selection: The EMRMR Approach

Most traditional IP networks face serious security and management challenges due to their rapid increase in complexity. SDN resolves these issues by the separation of control and data planes, hence enabling programmability for centralized management with flexibility. On the other hand, its centralized architecture makes SDN very prone to DDoS attacks, hence necessitating the use of advanced and efficient IDSs. This study focuses on improving IDS performance in SDN environments through the integration of deep learning techniques and novel feature selection methods. This study presents an Enhanced Maximum Relevance Minimum Redundancy (EMRMR) approach that incorporates a Mutual Information Feature Selection (MIFS) strategy and a new Contextual Redundancy Coefficient Upweighting (CRCU) strategy to optimize feature selection for early attack detection. Experiments on the inSDN dataset showed that EMRMR achieved better precision, recall, F1-score, and accuracy compared to the state-of-the-art approaches, especially when fewer features are selected. These results highlight the efficiency of the proposed EMRMR approach in the selection of relevant features with minimal computational overhead, which enhances the real-time capability for IDS in SDN environments.

T1WI Radiomics Analysis of Anterior Scalene Muscle: A Preliminary Application in Neurogenic Thoracic Outlet Syndrome.

This study aimed to analyze the differences in radiomic features of the anterior scalene muscle and evaluate the diagnostic performance of MRI-based radiomics model for neurogenic thoracic outlet syndrome (NTOS). Imaging data of patients with NTOS who underwent preoperative brachial plexus magnetic resonance neurography were collected and were randomly divided into training and test groups. The anterior scalene muscle area was sliced in the T1WI sequence as the region of interest for the extraction of radiomics features. The most significant features were identified using feature selection and dimensionality-reduction methods. Various machine learning algorithms were applied to construct regression models. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC). Totally, 267 radiomics features were extracted, of which 57 showed significant differences (P ≤ 0.05) between the abnormal and normal anterior scalene muscle groups. The least absolute shrinkage and selection operator regression model identified 13 optimal radiomic features with nonzero coefficients for constructing the model. In the training set, the AUROCs of diagnostic models built by different machine learning algorithms, ranked from highest to lowest, were as follows: support vector machine (SVM), 0.953; multilayer perception (MLP), 0.936; logistic regression (LR), 0.926; light gradient boosting machine (LightGBM), 0.906; and K-nearest neighbors (KNN), 0.813. In the testing set, the rankings were as follows: LR, 0.933; SVM, 0.886; KNN, 0.843; LightGBM, 0.824; and MLP, 0.706. NTOS is attributed to anterior scalene muscle abnormalities and exhibits distinct radiomic features. Integrating these features with machine learning can improve traditional manual image interpretation, offering further clarity in NTOS diagnosis.

Feature selection integrating Shapley values and mutual information in reinforcement learning: An application in the prediction of post-operative outcomes in patients with end-stage renal disease

Background:In predicting post-operative outcomes for patients with end-stage renal disease, our study faced challenges related to class imbalance and a high-dimensional feature space. Therefore, with a focus on overcoming class imbalance and improving interpretability, we propose a novel feature selection approach using multi-agent reinforcement learning. Methods:We proposed a multi-agent feature selection model based on a comprehensive reward function that combines classification model performance, Shapley additive explanations values, and the mutual information. The definition of rewards in reinforcement learning is crucial for model convergence and performance improvement. Initially, we set a deterministic reward based on the mutual information between variables and the target class, selecting variables that are highly dependent on the class, thus accelerating convergence. We then prioritised variables that influence the minority class on a sample basis and introduced a dynamic reward distribution strategy using shapley additive explanations values to improve interpretability and solve the class imbalance problem. Results:Involving the integration of electronic medical records, anesthesia records, operating room vital signs, and pre-operative anesthesia evaluations, our approach effectively mitigated class imbalance and demonstrated superior performance in ablation analysis. Our model achieved a 16% increase in the minority class F1 score and an 8.2% increase in the overall F1 score compared to the baseline model without feature selection. Conclusion:This study contributes important research findings that show that the multi-agent-based feature selection method can be a promising approach for solving the class imbalance problem.

Enhancing mortality prediction in patients with spontaneous intracerebral hemorrhage: Radiomics and supervised machine learning on non-contrast computed tomography

PurposeThis study aims to develop a Radiomics-based Supervised Machine-Learning model to predict mortality in patients with spontaneous intracerebral hemorrhage (sICH). MethodsRetrospective analysis of a prospectively collected clinical registry of patients with sICH consecutively admitted at a single academic comprehensive stroke center between January-2016 and April-2018. We conducted an in-depth analysis of 105 radiomic features extracted from 105 patients. Following the identification and handling of missing values, radiomics values were scaled to 0-1 to train different classifiers. The sample was split into 80-20% training-test and validation cohort in a stratified fashion. Random Forest(RF), K-Nearest Neighbor(KNN), and Support Vector Machine(SVM) classifiers were evaluated, along with several feature selection methods and hyperparameter optimization strategies, to classify the binary outcome of mortality or survival during hospital admission. A tenfold stratified cross-validation method was used to train the models, and average metrics were calculated. ResultsRF, KNN, and SVM, with the "DropOut+SelectKBest" feature selection strategy and no hyperparameter optimization, demonstrated the best performances with the least number of radiomic features and the most simplified models, achieving a sensitivity range between 0.90 and 0.95 and AUC range from 0.97 to 1 on the validation dataset. Regarding the confusion matrix, the SVM model did not predict any false negative test (negative predicted value 1). ConclusionRadiomics-based Supervised Machine Learning models can predictmortality during admission in patients with sICH. SVM with the "DropOut+SelectKBest"feature selection strategy and no hyperparameter optimization was the best simplifiedmodel to detect mortality during admission in patients with sICH.

Travel time prediction for an intelligent transportation system based on a data-driven feature selection method considering temporal correlation

Travel-time prediction is a critical component of Intelligent Transportation Systems (ITS), offering vital information for tasks such as accident detection, congestion management, and traffic flow optimisation. Accurate predictions are highly dependent on the selection of relevant features. In this study, a two-stage methodology is proposed which consists of two layers of Optimisation Algorithm and one Data-Driven method (OA2DD) to enhance the accuracy and efficiency of travel-time prediction. The first stage involves an offline process where interconnected optimisation algorithms are employed to identify the optimal set of features and determine the most effective machine learning model architecture. In the second stage, the real-time process utilises the optimised model to predict travel times using new data from previously unseen parts of the dataset. The proposed OA2DD method was applied to a case study on the M50 motorway in Dublin. Results show that OA2DD improves the convergence curve and reduces the number of selected features by up to 50%, leading to a 56% reduction in computational costs. Furthermore, using the selected features from OA2DD, reduced the prediction error by up to 29% compared to the full feature set and other feature selection methods, demonstrating the method's effectiveness and robustness.

Enhancing breast cancer diagnosis: a comparative analysis of feature selection techniques

Breast cancer is a significant contributor to female mortality, emphasizing the importance of early detection. Predicting breast cancer accurately remains a complex challenge within medical data analysis. Machine learning (ML) algorithms offer valuable assistance in decision-making and diagnosis using medical data. Numerous research studies highlight the effectiveness of ML techniques in improving breast cancer prediction. Feature selection plays a pivotal role in data preprocessing, eliminating irrelevant and redundant features to minimize feature count and improve classification accuracy. This study focuses on optimizing breast cancer diagnostics through feature selection methods, specifically genetic algorithms (GA) and particle swarm optimization (PSO). The research involves a comparative analysis of these methods and the application of a diverse set of ML classification techniques, including logistic regression (LR), support vector machine (SVM), decision tree (DT), and ensemble methods like random forest (RF), AdaBoost, and gradient boosting (GB), using a breast cancer dataset. The models' performance is subsequently evaluated using various performance metrics. The experimental findings illustrate that PSO achieved the highest average accuracy, reaching 99.6% when applied to AdaBoost, while GA attained an accuracy rate of 99.5% when employed with both AdaBoost and RF.

Radiomics based Machine Learning Models for Classification of Prostate Cancer Grade Groups from Multi Parametric MRI Images

Abstract Purpose: This study aimed to investigate the performance of multiparametric magnetic resonance imaging (mpMRI) radiomic feature-based machine learning (ML) models in classifying the Gleason grade group (GG) of prostate cancer. Methods: In this retrospective study, a total of 203 patients with histopathologically confirmed prostate cancer who underwent mpMRI before prostate biopsy were included. After manual segmentation, radiomic features (RFs) were extracted from T2-weighted, apparent diffusion coefficient, and high b-value diffusion-weighted magnetic resonance imaging (DWMRI). Patients were split into training sets and testing sets according to a ratio of 8:2. A pipeline considering combinations of two feature selection (FS) methods and six ML classifiers was developed and evaluated. The performance of models was assessed using the accuracy, sensitivity, precision, F1-measure, and the area under curve (AUC). Results: On high b-value DWMRI-derived features, a combination of FS method recursive feature elimination (RFE) and classifier random forest achieved the highest performance for classification of prostate cancer into five GGs, with 97.0% accuracy, 98.0% sensitivity, 98.0% precision, and 97.0% F1-measure. The method also achieved an average AUC for GG of 98%. Conclusion: Preoperative mpMRI radiomic analysis based on ML, as a noninvasive approach, showed good performance for classification of prostate cancer into five GGs. Advances in Knowledge: Herein, radiomic models based on preoperative mpMRI and ML were developed to classify prostate cancer into 5 GGs. Our study provides evidence that analysis of quantitative RFs extracted from high b-value DWMRI images based on a combination of FS method RFE and classifier random forest can be applied for multiclass grading of prostate cancer with an accuracy of 97.0%.

Multi-domain analysis of ultra-short-term HRV for breathing pattern classification in wearable health devices

This research paper introduces an innovative approach to classify heart rate variability (HRV) time series into paced and spontaneous breathing patterns to reflect changes in the autonomic nervous system. This type of classification is beneficial in wearable devices for stress/relaxation level detection and in deciding therapeutic interventions. The “Multi-Domain Approach” methodology integrates three different techniques: standard HRV features, fuzzy recurrence plot (FRP)-based FRP_GLCM, and empirical mode decomposition-based IMF_FRP_GLCM. The study concentrates on analyzing HRV time series within shorter data segments, aligning with the requirements of contemporary wearable health devices and biofeedback systems. HRV data collected during spontaneous and slow-paced breathing were analyzed across data segments of 5, 4, 3, 2, and 1 min, incorporating feature selection and reduction methods. Results demonstrated that standard HRV features yielded optimal performance for 5-min segments, achieving an average accuracy of 90%. Interestingly, IMF_FRP features achieved comparable accuracy even for 1-min segments. As segment duration decreased, standard HRV feature accuracy declined while IMF_FRP accuracy stayed intact, eventually matching 5-min segment accuracy levels. The study underscores the surging demand for shorter data segment HRV analysis, driven by advancements in wearable smart watches technology and mobile applications for monitoring health and managing stress.

Intrusion Detection in IoT using Gaussian Fuzzy Mutual Information-based Feature Selection

The proliferation of Internet of Things (IoT) devices has revolutionized various sectors by enabling real-time monitoring, data collection, and intelligent decision-making. However, the massive volume of data generated by these devices presents significant challenges for data processing and analysis. Intrusion Detection Systems (IDS) for IoT require efficient and accurate identification of malicious activities amidst vast amounts of data. Feature selection is a critical step in this process, aiming to identify the most relevant features that contribute to accurate intrusion detection, thus reducing computational complexity and improving model performance. Traditional Mutual Information-based Feature Selection (MIFS) methods face challenges when applied to IoT data due to their inherent noise, uncertainty, and imprecision. This study introduces a novel Fuzzy Mutual Information-based Feature Selection (Fuzzy-MIFS) method that integrates fuzzy logic with Gaussian membership functions to address these challenges. The proposed method enhances the robustness and effectiveness of the feature selection process, resulting in improved accuracy and efficiency of IDSs in IoT environments. Experimental results demonstrate that the Fuzzy-MIFS method consistently outperformed existing feature selection techniques across various neural network models, such as CNN, LSTM, and DBN, showcasing its superior performance in handling the complexities of IoT data. The results show that Fuzzy-MIFS increased the accuracy from 0.962 to 0.986 for CNN, from 0.96 to 0.968 for LSTM, and from 0.96 to 0.97 for DBN.

Machine learning-based performance prediction for energy storage medium-deep borehole ground source heat pump systems

With the rapid development of artificial intelligence, data-driven prediction models play an important role in energy demand forecasting and performance prediction. This study established performance prediction models for medium-deep borehole ground source heat pump (MDB-GSHP) systems to predict the coefficient of performance (COP) of GSHP, utilizing back propagation neural network (BPNN), extreme learning machine (ELM), support vector machine (SVM), and particle swarm optimization-support vector machine (PSO-SVM) method that optimizes the penalty parameter (C), insensitive loss parameter (ε), and kernel parameter (γ) of SVM using PSO. Operational data were collected through field experiments, and typical parameters were selected as features to establish the feature set. The feature set was then optimized using Pearson correlation analysis and recursive feature elimination (RFE), resulting in the creation of five distinct feature sets. Error analysis and trend evaluation results indicate that the combined feature set (FS5), which incorporates the advantages of multiple feature selection methods, demonstrates the best performance. The PSO-SVM model exhibits outstanding performance under various input conditions, achieving an accuracy of over 90% within the error range. When using FS5, the PSO-SVM model achieves a mean absolute percentage error (MAPE) of 0.037, mean absolute error (MAE) of 0.103, root mean square error (RMSE) of 0.125, coefficient of multiple determinations (R2) of 0.970, and explained variance score (EVS) of 0.967, showcasing excellent predictive performance. The research results can provide a basis for the formulation of storage-related operational parameters and the optimization of system operation for energy storage in MDB-GSHP heating systems.

Enhancing Leaf Area Index Estimation in Southern Xinjiang Fruit Trees: A Competitive Adaptive Reweighted Sampling-Successive Projections Algorithm and Three-Band Index Approach with Fractional-Order Differentiation

The Leaf Area Index (LAI) is a key indicator for assessing fruit tree growth and productivity, and accurate estimation using hyperspectral technology is essential for monitoring plant health. This study aimed to improve LAI estimation accuracy in apricot, jujube, and walnut trees in Xinjiang, China. Canopy hyperspectral data were processed using fractional-order differentiation (FOD) from 0 to 2.0 orders to extract spectral features. Three feature selection methods—Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), and their combination (CARS-SPA)—were applied to identify sensitive spectral bands. Various band combinations were used to construct three-band indices (TBIs) for optimal LAI estimation. Random forest (RF) models were developed and validated for LAI prediction. The results showed that (1) the reflectance spectra of jujube and walnut trees were similar, while apricot spectra differed. (2) The correlation between fractional-order differential spectra and LAI was highest at orders 1.4 and 1.7, outperforming integer-order spectra. (3) The CARS-SPA selected a smaller set of feature bands in the 1100~2500 nm, reducing collinearity and improving spectral index construction. (4) The RF model using TBI4 demonstrated high R², low RMSE, and an RPD value > 2, indicating optimal prediction accuracy. This approach holds promise for hyperspectral LAI monitoring in fruit trees.

Fuzzy multi-neighborhood entropy-based interactive feature selection for unsupervised outlier detection

Unsupervised feature selection is one of the important techniques for unsupervised knowledge discovery, which aims to reduce the dimensionality of conditional feature sets as much as possible to improve the efficiency and accuracy of the algorithm. However, existing methods have the following two challenges: (1) They are mainly applicable to select numerical or nominal features and cannot effectively select heterogeneous features; (2) The relevance and redundancy are primarily considered to construct feature evaluation indexes, ignoring the interaction information of heterogeneous features. To solve the challenges mentioned above, this paper proposes an unsupervised heterogeneous feature selection method based on fuzzy multi-neighborhood entropy, which also considers the multi-correlation of features to select heterogeneous features. First, the fuzzy multi-neighborhood granule is constructed by considering the distribution characteristics of the data. Then, the concept of fuzzy entropy is introduced to define the fuzzy multi-neighborhood entropy and its associated uncertainty measures, and the relationship between them is discussed. Next, the relevance, redundancy, and interactivity among attributes are defined, and the idea of maximum relevance-minimum redundancy-maximum interactivity is used to construct the evaluation indexes of heterogeneous features. Finally, experiments are conducted on several publicly unbalanced datasets, and the results are in comparison with existing algorithms. The experimental results show that the proposed algorithm can select fewer heterogeneous features to improve the efficiency of outlier detection tasks. The code is publicly available online at https://github.com/BELLoney/MNIFS.

Hybrid feature selection of microarray prostate cancer diagnostic system

DNA microarray prostate cancer diagnosis systems are widely used, and hybrid feature selection methods are applied to select optimal features to address the high dimensionality of the dataset. This work proposes a new hybrid feature selection method, namely the relief-F (RF)-genetic algorithm (GA) with support vector machine (SVM) classification method. The aim is to evaluate the performance of the proposed method in terms of accuracy, computation time, and the number of selected features. The method is implemented using Python in PyCharm and is evaluated on a DNA microarray prostate cancer. The outcome of this work is a performance comparison table for the proposed methods on the dataset. The performance of GA, particle swarm optimization (PSO), and whale optimization algorithm (WOA) is compared in terms of accuracy, computation time, and the number of selected features. Results show that GA has the highest average accuracy (91.17%) compared to PSO (90.52%) and WOA (85.74%). GA outperforms PSO and WOA due to its superior convergence properties and better alignment with complex problems.

Feature selection in high-dimensional classification via an adaptive multifactor evolutionary algorithm with local search

As datasets grow in dimension and sample size, feature selection becomes increasingly important in machine learning. Features are often associated with multiple tasks, so adopting a multi-task optimization framework in feature selection can improve its classification performance. Multifactor optimization provides a powerful evolutionary multi-tasking paradigm capable of simultaneously handling multiple related optimization tasks. Taking inspiration from these, this article proposes a parameter adaptive multifactor feature selection algorithm (AMFEA). To help the algorithm escape from local optima, AMFEA uses a local search strategy to assist the algorithm in finding the global optimum. In addition, AMFEA has designed an adaptive knowledge transfer parameter matrix that dynamically adjusts parameter sizes based on the population’s fitness to control the frequency of knowledge transfer between tasks. This effectively transfers knowledge between different tasks and helps the algorithm converge quickly. Experimental results on 18 high-dimensional datasets show that AMFEA significantly improves classification accuracy compared with evolutionary algorithms and traditional feature selection methods.

A New COVID-19 Patient Detection Strategy Based on Hidden Naïve Bayes Classifier

COVID-19 is a universal infectious disease recognized first by people with influenza and bacterial pneumonia symptoms in Wuhan, Hubei Province. Currently, a new mutated disease has the same symptoms as COVID-19 and influenza and causes dangerous infections in the body. Due to the fact that these two diseases share some diagnostic features and symptoms in common with one another, healthcare workforces require aid and support in predicting patients' conditions. This was done by using machine learning methods in diagnosis. From this point, this paper proposes a diagnostic model to detect patients' symptoms and classify them into one of five disease groups, utilizing Neighborhood Component Analysis (NCA) as a feature selection method and the Hidden Naïve Bayes (HNB) method as a multiclass classifier. This paper suggests the model consists of two significant phases: the pre-processing phase (cleaning, normalization, and discretization) and the classification phase. Conducting the COVID-19 dataset, the experimental findings showed that the suggested multi-class model had 89% accuracy for disease diagnoses. Furthermore, according to the patient’s symptoms, the proposed classification model led to a good diagnosis for the mutated COVID-19 disease.

Plant-level prediction of potato yield using machine learning and unmanned aerial vehicle (UAV) multispectral imagery

This study presents a new method for predicting the underground yield of potato at the plant level, using two key approaches: (1) identifying the critical variables for yield prediction based on plant height and vegetation index (VI) maps derived from unmanned aerial vehicle (UAV) imagery; (2) evaluating the accuracy of predictions for fresh tuber weight (FTW), number of tubers (NMT), and fresh weight per tuber (FWT), using various machine learning (ML) algorithms. During the growing season of 2022, high-resolution red, green, and blue light and multispectral images were collected weekly using a UAV. In total, 648 variables, including first- and second-order statistical parameters, were extracted from the images. Five feature-selection algorithms were used to identify the key variables influencing the predictions of FTW, NMT, and FWT. Furthermore, ML models, including random forest (RF), ridge regression, and support vector machines, were employed to refine the variable sets for ensuring stable yield component predictions. The results highlighted the importance of considering first- and second-order statistical parameters derived from plant height and VI. Second-order statistics were crucial for predicting the FTW and FWT. The RF model demonstrated high prediction accuracy, with R2 values of 0.57, 0.45, and 0.49 for FTW, NMT, and FWT, respectively, using the best feature selection method. Thus, leveraging RGB and multispectral imagery data recorded that 1.5–2 months before harvest can significantly enhance yield predictions conducted using ML models. The proposed methodology can help farmers growing potatoes or other crops optimize cultivation and predict the yield.