HighlightsWhat are the main findings?The improved ROF-LightGBM model with optimized base classifiers achieves higher classification accuracy and computational efficiency than conventional rotation forest, random forest and LightGBM models. It also possesses better robustness under limited training samples and noisy data interference.In this study, minimum noise fraction (MNF) is introduced to construct the rotation matrix for the improved ROF-LightGBM framework. Compared with PCA and ICA, MNF is more suitable for lithological classification tasks, which further boosts the overall performance of the modified model.What are the implications of the main findings?This study proposes an optimized ROF-LightGBM (MNF) framework to address the low computational efficiency of conventional ROF algorithms when processing high-dimensional remote sensing data.The method realizes effective lithological recognition in bedrock outcrops, offering solid technical support for high-precision geological investigation and mapping.Hyperspectral remote sensing data provides distinct advantages for lithological classification in bedrock-exposed areas. Despite the superior performance of ensemble learning methods (e.g., Rotation Forest, ROF) in big data classification, their application in high-dimensional hyperspectral data is restricted by high training costs. To address this limitation and improve classification accuracy, this study proposes an optimized ROF-LightGBM ensemble algorithm integrated with minimum noise fraction (MNF) for rotation matrix construction. Experimental validation was conducted using ZY1-02D hyperspectral data for lithological mapping in the bedrock-exposed Xitieshan area, involving ROF-LightGBM parameter optimization (L × T, bootstrap) and comparative experiments with multiple machine learning models. The results demonstrate the following: under the same number of decision trees (T = 100), the ROF-LightGBM (PCA, L × T = 4 × 25) with optimized base classifier outperforms random forest (RF), LightGBM, and traditional ROF (L = 100) models in classification accuracy, achieving 74.28% accuracy, 6.54% higher than RF and 1.53% higher than LightGBM. More notably, it boasts exceptional efficiency, with a training time of only 4.86 s (nearly 37 times shorter than traditional ROF), while maintaining minimal accuracy loss (an only 1.19% decrease). Additionally, the ROF-LightGBM (MNF) model, which adopts MNF for rotation matrix construction, further enhances performance. Compared with the PCA-based ROF-LightGBM, it achieves an 82.17% classification accuracy (a 7.89% increase) and its kappa coefficient reaches 0.81, fully verifying the model’s superiority in accuracy and efficiency.

Semi-Supervised Rotation Forest

Breast cancer is a widespread disease involving abnormal (uncontrolled) growth of breast tissue cells along with the formation of a tumor and metastasis. Breast cancer cases occur mostly among women. Early detection and regular screening have significantly improved survival rates. This research classifies breast cancer and non-breast cancer cases using machine learning algorithms based on the Breast Cancer Coimbra dataset by optimizing the classifier performance and feature selection methodology. In addition, this research identifies the influential features responsible for BC classification by using diverse counterfactual explanations. The Rotation Forest classifier algorithm is used to classify breast cancer and non-breast cancer cases. The hyperparameters of this algorithm are optimized using the Optuna optimizer. Three wrapper-based feature selection techniques (Sequential Forward Selection, Sequential Backward Selection, and Exhaustive Feature Selection) are used to select the most relevant features. An ensemble environment is also created using the best feature subsets of these methods, incorporating both soft and hard voting strategies. Experimental results show that the hard voting strategy achieves an accuracy of 85.71%, F1-score of 83.87%, precision of 92.85%, and recall of 76.47%. In contrast, the soft voting strategy obtains an accuracy of 80.00%, F1-score of 77.42%, precision of 85.71%, and recall of 70.59%. These findings demonstrate that hard voting achieves noticeably better performance. The misclassification outcomes of both strategies are explored using Diverse Counterfactual Explanations, revealing that BMI and Glucose values are most influential in predicting correct classes, whereas the HOMA, Adiponectin, and Resistin values have little influence.

The escalating risks posed by earthquakes necessitate advanced methods in building design, seismic response prediction, and evaluation. Despite the presence of conventional methods like inverse and forward approaches, which include state-space models and finite element models (FEMs), limitations persist in terms of computational demands and data constraints. This article employs four machine learning (ML) models including Decision Tree (DT), Decision Stump, Rotation Forest, and Random Forest to predict the seismic response of an existing building in Busan city, South Korea. The dataset with a total of 780,000 data points, in-situ observations obtained through a Structural Health Monitoring (SHM) system via real-world signals from accelerometers and GPS displacement gauges placed in a building. Time Series Data Splitting method, which is applied for cross-validation, ensures that the training set always precedes the test set in chronological order, preserving the temporal structure inherent in time series datasets. The use of differential Continuous Wavelet Transform (CWT) analysis is instrumental in delineating the strengths and weaknesses of each model. The results demonstrate that the DT model closely aligns with the actual displacement trends, showcasing its efficacy in seismic response prediction. This approach not only enhances the accuracy of identifying and detecting building responses but also offers a resource-efficient alternative to traditional methods.

Groundwater pollution presents a serious concern in arid and semiarid regions, where water resources are already limited. In such contexts, reliable and efficient methods for assessing groundwater vulnerability are critical. Without adequate knowledge of the vulnerability, groundwater is at greater risk of severe contamination. This not only threatens the availability of clean water but also demands significant time and financial resources for remediation and restoration. Modelling groundwater vulnerability is even more demanding and complex in data-scarce regions. Consequently, this study investigates and predicts spatial variations in groundwater quality and vulnerability within a data-scarce area by applying efficient machine learning methods that compensate for the limited availability of quality groundwater data. Supportive machine learning approaches such as bagged adaptive boosting (BAB), averaged neural network (avNNet), heteroscedastic discriminant analysis (HAD), rotation forest (RotationF), and an ensemble method were applied to assess groundwater vulnerability using k-fold cross-validation. The results demonstrate that the BAB model achieved the best performance, with both accuracy and precision exceeding 85%. Furthermore, the stacking ensemble approach, specifically the BAB model combination, increased precision by 4% and reduced false alarms by 6%. The most influential variables affecting groundwater quality include groundwater depth, precipitation, proximity to waterways and roads, topographic humidity, and the percentage of fine-grain material. The results also show that variability in the data significantly impacts the modelling performance.

ABSTRACT The Ministry of Agriculture-Jihad (MAJ) and the Iranian Space Agency (ISA) aim to accurately estimate the cultivated area of strategic crops and evaluate their annual yield through meticulous crop mapping. However, Iran lacks a comprehensive, integrated approach using remote sensing and machine learning for this purpose. This study addressed this gap by developing a versatile, user-friendly crop mapping framework for Iran, utilizing Landsat-8 time series data and classical machine learning algorithms. Marvdasht in Fars province was selected as the pilot area due to its high diversity of agricultural crop types and its status as a significant agricultural hub in Iran. Furthermore, the most widely used and flexible methods available in crop mapping studies such as decision tree (DT), random forest (RF), rotation forest (RoF), support vector machine (SVM), and dynamic time warping (DTW) were used in this study. The results showed that the DTW and RF methods outperformed others, achieving approximately 96% accuracy and improving overall accuracy by 8% in creating the crop map for the pilot area. Additionally, this study demonstrated the effectiveness of Landsat-8 bands 2 to 5 along with the normalized difference vegetation index (NDVI) in reliably identifying all crops in the region. The proposed framework shows promise for significantly advancing crop mapping practices in Iran.

Envelope rotation forest: A novel ensemble learning method for classification

Some machine learning (ML) methods used for variable importance estimation are considered. The paper provides an overview of several tree-based classification methods (CART, AdaBoost, Gradient Boosted Trees, Random Forest, Extra Trees, Rotation Forest) and compares peculiarities of their application to biomedical data analysis. Additionally, possible adaptations of the algorithms for processing datasets with imbalanced classes are proposed.

Abstract Global warming caused by climate change causes some problems in agricultural production. One of these problems is the increase in various pest populations. This increase poses a serious threat to agricultural products and significantly negatively affects productivity and quality. Insecticides are commonly used to combat pests. However, most of the time, farmers' lack of knowledge in recognizing pests and understanding their effects results in incorrect and excessive spray applications. While excessive use of insecticides harms human health and environmental pollution, it also increases production costs, causes changes in the genetic structures of pests, causing them to become more resistant, and makes agricultural control difficult. Therefore, early detection of pests and their damage to the plant is extremely important. This study aims to develop an accurate and efficient method to detect damage caused by the tomato leaf miner, Tuta absoluta, on tomato leaves. A dataset comprising healthy and damaged tomato leaves was created. Using a hybrid approach, features were extracted through Convolutional Neural Networks (CNNs) with transfer learning and classified using traditional machine learning techniques. Among the methods evaluated, SVM-Linear achieved the highest accuracy with 97.83%, outperforming other classifiers such as Random Forest with 96.14%, Rotation Forest with 95.89%, and SVM-RBF with 90.70%. These results highlight the potential of combining deep learning-based feature extraction with conventional machine learning for early pest detection. This approach offers a practical solution to reduce the misuse of insecticides and improve pest management strategies, contributing to sustainable agriculture.

Abstract. Poplars (Populus sp.), a tree that grows rapidly species, are significant as industrial forest products. The delineation and monitoring of poplar cultivated areas are invaluable for decision-making processes. With the remote sensing technology, accurate detection of poplar planted areas could be determined much faster, more economically, and with minimum labor requirements. The objective of this research is to create a map of poplar plantations in the Sakarya region of Turkey utilizing Worldview-3 satellite imagery. Object-based image analysis (OBIA) through the application of the multi-resolution segmentation method (MRS) was employed to generate image segments, and then three prevailing machine learning algorithms, namely Support Vector Machine (SVM), Random Forest (RF) and Rotation Forest (RotFor) were implemented to produce LULC maps of the study area including 11 landscape features. The most effective and contributing object features that assure high separability between landscape features were determined using a filter-based Chi-square algorithm for the prediction models constructed with SVM, RF, and RotFor classifiers. Results revealed that the SVM classifier achieved the highest overall accuracy (91.73%) with 38 features out of 88 features, about 3% improvement compared to the other algorithms. According to the SHAP analysis, the IHS feature was the most effective one in the constructed RF model, followed by the CI (red edge), NDVI-1 and NDVI-2 vegetation indices.

Majuli is the world’s largest inhabited river island and is highly prone to flood hazards, resulting in significant damage to houses and agriculturally based livelihoods. Considering its cultural heritage and unique landscape, it is necessary to prepare a flood susceptibility map (FSM) to reduce the annual damage. Therefore, the primary aim of this research is to prepare and improve the precision of FSM using microwave satellite images and six robust ensemble machine learning models. In the three main stages of FSM, each stage contributes to achieving optimal accuracy. In the first stage, a threshold-based flood inventory map has been prepared from six years of multi-temporal SAR images. In the second stage, preliminarily seventeen flood conditioning variables such as elevation, slope, profile curvature, terrain ruggedness index, topographic wetness index, distance from streams, rainfall, land cover land use, normalized difference vegetation index, distance from road, geomorphology and lithology have been prepared, but after utilizing the Boruta algorithm and multicollinearity analysis, twelve key flood-influencing variables have been selected for flood modelling. In the final stage, six robust ensemble machine learning models namely random forest, rotation forest, stochastic gradient boosting, boosted regression tree, deep boost and logit boost have been applied and subsequently compared to determine the best model. The performance of the models is evaluated with various statistical measurements, including the area under curve (AUC), sensitivity, specificity, kappa index and overall accuracy values. The results revealed that the random forest model outperformed the other models in terms of model fitness (AUC = 1) and predictive capability (0.99). Additionally, the very highly vulnerable pixels of the FSM are validated with the twenty flood locations from the field surveys, showing that the accuracy of the FSM is 100%. The FSM indicates that around 50% of the study region has a very high and high susceptibility to future flood occurrences.

Multi-classification prediction of PM2.5 concentration based on improved adaptive boosting rotation forest

Land degradation (LD) is the decline in a land's functional capacity and productive potential, which includes various anthropogenic and natural drivers. This study focuses on three primary manifestations of LD including soil erosion, landslides, and rockfalls, which are the most prevalent in the Shaqlawa district. A set of 22 LD conditioning factors, encompassing curvature, lithology, aspect, river density, soil type, lineament density, river distance, elevation, road distance, length slope (LS), land use land cover (LULC), stream power index (SPI), valley depth, profile curvature, slope, solar radiation, road density, lineament distance, rainfall, topographic wetness index (TWI), plan curvature, and normalized difference vegetation index (NDVI), were integrated into the analysis. Variance inflation factors (VIF) and tolerance (TOL) values from linear regression indicate that most LD factors have acceptable levels of multicollinearity. The Information Gain Ratio (IGR) identified key variables TWI, NDVI, and lithology-as pivotal factors for predicting LD. Additionally, the study evaluated degradation factors using various machine learning (ML) algorithms, including random forest (RF), Naive Bayes, logistic regression, rotation forest, forest penalized attributes (FPA), and Fisher's Linear discriminant analysis (FLDA). This facilitated categorizing the study area into five susceptibility categories. The FLDA model categorized the highest area under very high degradation risk at 26.72%, emphasizing the varied insights each algorithm brought to characterizing the degradation risk. Additionally, the receiver operating characteristic curves (ROC) were employed for model validation, identifying RF as the most successful model in the training dataset with an area under the curve (AUC) of 0.882, while FLDA outperformed in the testing dataset with an AUC of 0.883. The identified LD-prone areas will help land-use planners and emergency management officials apply effective mitigation strategies for similar terrains.

The present study investigates the effectiveness of various landslide susceptibility machine learning (ML) models at multiple spatial resolutions. Using various conditioning factors, including topography, hydrology, and human influences, the study analyzed the predictive power of single, integrated, and comparative ML models. Alternating Decision Tree (ADT), Forest by Penalizing Attributes (FPA), and Random Forest (RAF), and integrated approaches such as Rotation Forest (RF) and Random Subspace (RS) that were based on ADT and FPA were utilized for the study. All models were trained and validated using a ten-fold cross-validation technique and evaluated by various statistical metrics, across resolutions from 12.5 m to 200 m. Shenmu City in China was chosen as an ideal test site to evaluate the developed methodology. The study reveals that finer spatial resolutions significantly enhance the accuracy of landslide predictions and that integrated models have superior performance over single models. The Frequency Ratio method identified elevation, slope, and hydrological factors as the key predictors concerning landslide occurrences. According to the results the RS-ADT model at a 12.5 m resolution achieved the highest evaluation accuracy (ROC = 0.907). The research highlights the possibility of combining multiple modeling techniques and high-resolution data to improve the predictive accuracy of landslide susceptibility models.

Accurate prediction of daily river flow (Qt) remains a challenging yet essential task in hydrological modeling, particularly crucial for flood mitigation and water resource management. This study introduces an advanced M5 Prime (M5P) predictive model designed to estimate Qt as well as one- and two-day-ahead river flow forecasts (i.e. Qt+1 and Qt+2). The predictive performance of M5P ensembles incorporating Bootstrap Aggregation (BA), Disjoint Aggregating (DA), Additive Regression (AR), Vote (V), Iterative classifier optimizer (ICO), Random Subspace (RS), and Rotation Forest (ROF) were comprehensively evaluated. The proposed models were applied to a case study data in Tuolumne County, US, using a dataset comprising measured precipitation (Pt), evaporation (Et), and Qt. A wide range of input scenarios were explored for predicting Qt, Qt+1, and Qt+2. Results indicate that Pt and Qt significantly influence prediction accuracy. Notably, relying solely on the most correlated variable (e.g., Qt-1) does not guarantee robust prediction of Qt. However, extending the forecast horizon mitigates the influence of low-correlation input variables on model accuracy. Performance metrics indicate that the DA-M5P model achieves superior results, with Nash-Sutcliff Efficiency of 0.916 and root mean square error of 23 m3/s, followed by ROF-M5P, BA-M5P, AR-M5P, AR-M5P, RS-M5P, V-M5P, ICO-M5P, and the standalone M5P model. The ensemble M5P modeling framework enhanced the predictive capability of the stand-alone M5P algorithm by 1.2 %–22.6 %, underscoring its efficacy and potential for advancing hydrological forecasting.

Abstract Gully erosion is a significant natural hazard and a form of soil erosion. This research aims to predict gully formation in the Kalshour basin, Sabzevar, Iran. Employing the Information Gain Ratio (IGR) index, we identified 13 key factors out of 22 for modeling, with elevation emerging as the most influential factor in gully formation. The study evaluated the performance of individual machine learning algorithms and ensemble algorithms, including the Functional Tree (FT) as the main classifier, Bagging (Bagg), AdaBoost (Ada), Rotation Forest (RoF), and Random Subspace (RSS). Using a data set of 400 gully and non‐gully points obtained through field investigations (70% for training and 30% for testing), the RoF model achieved an area under the curev (AUC) value of 0.99, indicating its high predictive ability for gully‐susceptible areas. Other algorithms also performed well (Ada: 0.90, FT: 0.92, RSS: 0.94, Bagg: 0.95). However, the RoF algorithm with the functional tree as the main classifier (RoF_FT) demonstrated the highest ability in gully classification and susceptibility mapping, enhancing the functional tree's performance. In addition to AUC, the RoF_FT model achieved an F1 score of 0.89 and an MCC of 0.78 on the validation set, indicating a high balance between precision and recall, and a strong correlation between predicted and actual classes, respectively. Similarly, other models showed robust performance with high F1 scores and MCC values, but the RoF_FT model consistently outperformed them, underscoring its robustness and reliability. The resulting gully erosion‐susceptibility map can be valuable for decision‐makers and local managers in soil conservation and minimizing damages. Implementing proactive measures based on these findings can contribute to sustainable land management practices in the Kalshour basin.Recommendations Gully erosion threat: Gully erosion poses a significant threat to soil, with far‐reaching environmental consequences. Predictive modeling: This research focuses on predicting gully formation in the Kalshour basin, Sabzevar, Iran, using advanced machine learning algorithms. Key findings for decision‐makers: The study evaluates the performance of various machine learning algorithms and ensemble algorithms, with the Functional Tree serving as the main classifier. This not only enhances our ability to predict gully formation but also provides a valuable tool for decision‐makers and local managers in soil conservation. Impact on sustainable land management: By offering a gully erosion‐susceptibility map, the research empowers decision‐makers to implement proactive measures, minimizing damage and contributing to sustainable land management practices. Interdisciplinary approach: The study's combination of geospatial analysis, machine learning, and soil conservation aligns with the journal's mission to advance understanding in environmental modeling.

Purpose: The per capita expenditure data in Jambi Province, Indonesia have been plagued with severe multicollinearity problems. To address the issue, this study developed an effective small area estimation (SAE) method, which is essential for formulating comprehensive regional development policies in Jambi Province. By modifying the mixed effects random forest (MERF) method, we introduced PCA-MERF (which applies principal component analysis prior to MERF) and MERoF (which replaces the standard random forest with rotation forest) to handle multicollinearity more effectively. Data from the National Socioeconomic Survey (Susenas) in March 2021 and Village Potential (PODES) in 2021 were utilized. The methods were evaluated using metrics such as root mean square error (RMSE), relative root mean square error (RRMSE), coefficient of variation (CV), and their ability to capture random area effects. The random effect block (REB) bootstrap approach was employed to obtain MSE estimates for evaluating area-level estimate quality. Result: The results showed that MERoF outperformed both MERF and PCA-MERF, particularly in unit-level (village) estimation. Additionally, MERoF demonstrated superior capability in capturing variation between subdistricts compared to MERF and PCA-MERF. PCA-MERF performed better than MERF and MERoF at the area level (subdistrict). All three methods showed acceptable performance with RRMSE and CV values ranging between 8% and 10%, indicating precise and reliable predictions for per capita expenditure in small areas. These modifications to MERF prove effective and advantageous for small-area estimation in datasets with significant multicollinearity. Novelty: This research introduces a novel semi-parametric, tree-based SAE approach, enhancing the precision of per capita expenditure estimates and supporting more informative regional policy decisions, thus filling a gap in current SAE methodologies.

Parkinson's disease (PD) is a neurodegenerative disorder characterized by diverse symptoms, where accurate diagnosis remains challenging. Traditional clinical observation methods often result in misdiagnosis, highlighting the need for biomarker-based diagnostic approaches. This study utilizes ultraperformance liquid chromatography coupled to an electrospray ionization source and quadrupole time-of-flight untargeted metabolomics combined with biochemometrics to identify novel serum biomarkers for PD. Analyzing a Brazilian cohort of serum samples from 39 PD patients and 15 healthy controls, we identified 15 metabolites significantly associated with PD, with 11 reported as potential biomarkers for the first time. Key disrupted metabolic pathways include caffeine metabolism, arachidonic acid metabolism, and primary bile acid biosynthesis. Our machine learning model demonstrated high accuracy, with the Rotation Forest boosting model achieving 94.1% accuracy in distinguishing PD patients from controls. It is based on three new PD biomarkers (downregulated: 1-lyso-2-arachidonoyl-phosphatidate and hypoxanthine and upregulated: ferulic acid) and surpasses the general 80% diagnostic accuracy obtained from initial clinical evaluations conducted by specialists. Besides, this machine learning model based on a decision tree allowed for visual and easy interpretability of affected metabolites in PD patients. These findings could improve the detection and monitoring of PD, paving the way for more precise diagnostics and therapeutic interventions. Our research emphasizes the critical role of metabolomics and machine learning in advancing our understanding of the chemical profile of neurodegenerative diseases.

The article introduces datasets representing piston pump failures along with the experimental evaluation of various machine learning classification models. It starts with a detailed description of three classification datasets consisting of three different levels of valve plate damages and signals recorded from sensors used in classical hydraulic systems (pressure, temperature, flow). The obtained datasets consist of 100k (Failure 1), 30k (Failure 2) and 30k (Failure 3) samples and eight attributes. Then a broad range of classifiers are evaluated including three ensemble models based on decision trees: Random Forest, Gradient-Boosted Trees, and Rotation Forest, as well as the kNN algorithm and a neural network. The analysis showed that neural networks achieved the highest prediction accuracy, enabling a prediction accuracy level of 89%. The kNN algorithm ranked second, and tree-based algorithms performed 4% worse than the neural network. Next, the attribute importance analysis revealed that leak flow, pressure output, pressure of the leak line, and oil temperature are the most important parameters for accurate predictions. Additionally, the research includes a sensitivity analysis of the best classifier to verify the impact of sensor measurements or other noise indicators on the prediction model performance. The analysis indicates a 5% margin of measurement quality.

Nutrient extraction is related to stem diameter distribution, tissue concentration, and yield in an annually harvested Salix coppice