High spatial resolution remote sensing imagery encompasses multiple low-level features (e.g., spectral characteristics, patterns, shadows, and textures) and explicit high-level semantic features. High-resolution remote sensing scene classification (RSSC) is essential for environmental understanding and has gained significant attention in remote sensing image interpretation (Gu et al., 2019; Cheng et al., 2017, 2020). Scene classification involves categorizing images into predefined semantic classes based on scene-specific information, representing high-level abstractions of scene content. However, traditional classification methods are limited by their reliance on low-level feature analysis, which restricts their capacity to extract high-level semantic information, thereby often failing to meet the accuracy demands of RSSC. Currently, the proliferation of deep learning has spurred numerous methodologies for remote sensing scene image classification, which can be broadly categorized into three types: autoencoder-based, convolutional neural network (CNN)-based, and generative adversarial network (GAN)-based approaches (Cheng et al., 2017; Ma et al., 2019; Yu et al., 2020; Cheng et al., 2022). For instance, Du et al. (2017) proposed stacked convolutional denoising auto-encoders, which demonstrated superior classification performance compared to state-of-the-art unsupervised networks. Tang et al. (2021) proposed the attention consistent network (ACNet) based on Siamese networks and validated the method using three remote sensing scene datasets, demonstrating that the proposed method achieves good performance. Cheng et al. (2022) proposed perturbation-seeking generative adversarial networks (PSGANs) to improve the stability of sample generation. Recent advances have also incorporated attention mechanisms and multiscale feature fusion, with notable contributions including co-enhanced global-part integration approaches (Zhao et al., 2024a), gradient-guided multiscale focal attention networks (Zhao et al., 2024b), and multibranch fusion-based feature enhancement methods (Liu et al., 2019), which have demonstrated remarkable performance improvements.

Despite these extensive advances in RSSC, the challenges posed by limited labeled samples and class imbalance remain insufficiently addressed. Deep learning, as a data-driven paradigm, relies heavily on data availability and quality, computational resources, architectural innovations, and optimization techniques. Among these factors, dataset quality is a critical determinant of model performance, with class balance being a key quality indicator. When training datasets are imbalanced, classifiers tend to favor majority classes and often fail to correctly identify minority classes (Buda et al., 2018; Johnson and Khoshgoftaar, 2019; Leevy et al., 2018; Luque et al., 2019; Yessou et al., 2020; Thabtah et al., 2020). In extreme cases, particularly with noise-robust models, minority classes may even be completely ignored, leading to severe classification bias.

Approaches to addressing class imbalance in machine learning are typically divided into two categories: data-level methods and algorithm-level methods. Data-level methods aim to modify the distribution of training samples to improve the effectiveness of standard algorithms (Fernández et al., 2018). For instance, Liu and Tsoumakas (2020) used random under-sampling to rebalance class distribution. Among over-sampling techniques, the synthetic minority oversampling technique (SMOTE) has been recognized as one of the most influential data-level strategies (Douzas et al., 2018). SMOTE is fundamentally a clustering-based method designed to oversample one-dimensional vector data—such as spectral vectors in multispectral or hyperspectral images (Feng et al., 2019)—to alleviate overfitting. However, its effectiveness is limited when applied to two-dimensional or three-dimensional image data due to its tendency to introduce noise. In contrast, algorithm-level methods maintain the original data distribution and instead focus on adjusting training strategies or inference mechanisms. For example, Bria et al. (2020) proposed a two-stage deep learning framework to address severe class imbalance in small lesion detection. Ren et al. (2020) introduced an improved DeepLab V3+ for remote sensing image segmentation, incorporating a loss function tailored to sample distribution. In addition, few-shot learning approaches offer alternative perspectives on handling limited minority class samples. Recent studies by Deng et al. (2024) on masked second-order pooling for few-shot RSSC demonstrates the potential of meta-learning approaches. Fine-tuning strategies have also been shown to enhance performance under imbalanced conditions, with Guan et al. (2020) developing a random fine-tuning meta metric learning (RF-MML) model for aerial image classification, which demonstrated effectiveness in handling imbalanced data. Despite these advancements, the effectiveness of class imbalance mitigation strategies, particularly data augmentation approaches, in deep learning-based RSSC tasks remains systematically underexplored.

While high-resolution remote sensing images contain rich semantic information, their scene classification faces significant challenges from limited labeled samples and class imbalance. Although some studies have investigated class imbalance in remote sensing contexts (Guan et al., 2020), comprehensive research in this area remains limited. This study aims to address these gaps through a systematic investigation of class-imbalanced RSSC. The specific objectives of this study are:

• Evaluating the sensitivity of overall accuracy metrics in RSSC to class imbalance

• Investigating the integration of data augmentation methods with datasets and examining the feasibility of alleviating classification bias through data augmentation methods

• Analyzing the sensitivity of individual categories in RSSC to class imbalance

• Assessing the sensitivity of RSSC to different classification algorithms

Figure 1 illustrates the experimental workflow. First, a class-imbalanced high-resolution remote sensing scene image dataset is constructed. Second, comprehensive data augmentation methods are applied according to dataset characteristics, reducing the dataset's imbalance ratio (IR) by approximately one-seventh. Subsequently, the classifiers are trained and fine-tuned on both augmented and original datasets. Finally, model performance on overall and individual categories is evaluated, with visualization of results from selected study areas.

The main contributions of this paper include investigating the feasibility of alleviating the class-imbalance problems using data augmentation methods in RSSC and advancing research in this field. In addition, a class-imbalanced RSSC benchmark dataset is constructed, and deep learning methods with different parameters and structures are evaluated.

As for the organization of this article, Section 2 introduces the methodology and experimental design in detail. Sections 3 and 4 present and discuss the experimental results comprehensively. Finally, Section 5 summarizes the conclusions and outlines future research directions.

The relationship between training sample size and class imbalance sensitivity is fundamental to understanding classifier performance in remote sensing applications. The experimental results demonstrate that categories with fewer training samples experience disproportionately greater impact from class imbalance. Specifically, C4 (intersections), representing the minority class with the lowest proportion (0.037, 68/1,858) in the class-imbalanced dataset, exhibited significant classification deviation across all tested architectures. This phenomenon led to poor convergence during training and notable performance degradation for this category. In contrast, C8 (water), the majority class with the highest proportion (0.343, 638/1,858), achieved near-optimal performance due to its abundant training samples and relatively uniform spectral characteristics. Consequently, all models attained recall and precision values approaching 100% for this category.

Comparative analysis between models trained on class-imbalanced and class-balanced datasets reveals substantial improvements across all evaluation metrics, including overall accuracy, Kappa coefficient, recall, and precision. While the water body (C8) showed minimal performance changes due to its already optimal baseline, the intersections (C4) demonstrated remarkable improvements of 72% and 128% in precision and recall, respectively. This pattern confirms that data augmentation provides the greatest benefits for categories with the most limited training samples, effectively addressing the core challenge of class imbalance.

The relationship between architectural complexity and class imbalance robustness reveals counterintuitive patterns that challenge conventional assumptions. MobileNet-v2, despite its lightweight design with a model size approximately 1/18th that of ResNet101 (8.83 MB vs. 162 MB), achieved competitive performance on balanced datasets with only a 2.6% accuracy gap. However, under class-imbalanced conditions, MobileNet-v2 exhibited significant vulnerability, particularly in minority classes, due to its limited model capacity (Kamilaris and Prenafeta-Boldú, 2018). More surprisingly, ResNeXt101_32 × 32d, with model complexity approximately 11 times greater than ResNet101 (1,740 MB vs. 162 MB), demonstrated inferior performance on imbalanced datasets despite incorporating advanced architectural features such as grouped convolutions and residual connections. This counterintuitive result suggests that excessive model complexity may lead to overfitting on majority classes while failing to adequately capture minority class representations. The introduction of Transformer (ViT-B/16) provides additional insights, as it achieved superior overall performance and enhanced stability across imbalanced classes, demonstrating that attention-based mechanisms offer inherent advantages for handling class distribution skew.

RSSC inherently involves semantic complexity that extends beyond traditional class imbalance challenges. Individual scene images frequently encompass multiple ground objects with diverse semantic categories, with semantic classification typically determined by the category with the greatest likelihood (Dunne and Campbell, 1997; Liang et al., 2017). However, results analysis combined with existing literature reveals that semantic categorization becomes fundamentally ambiguous for many scene images due to the inherent complexity and overlap of ground objects. This observation suggests that multilabel classification frameworks may be more appropriate for remote sensing applications, particularly in heterogeneous urban environments.

Furthermore, data augmentation necessitates category-specific adjustments to mitigate the adverse effects of noisy data (Kim et al., 2003; Qi et al., 2018). For example, the cropping scale and position of scene images can influence their semantic category, potentially introducing noisy data through the application of cropping methods. The red-framed region (512^*512) in Figure 10, classified as an intersection, is segmented into four distinct scene areas (a, b, c, and d) at a 256^*256 scale. Areas a, b, and c can be categorized as road, while area d represents open-air venues, demonstrating a clear divergence in semantic categories across the two cropping scales. Moreover, a horizontal shift of approximately 128 pixels to the right for area “a” reclassifies it as open-air venues, further illustrating the impact of positional variance on semantic categorization. Additional data augmentation techniques, such as inappropriate rotation angles and excessive HSV disturbance amplitudes, may also contribute to the introduction of noisy data within the dataset.

While conventional augmentation methods prove effective, there remains significant potential for developing specialized techniques that better exploit the unique characteristics of remote sensing data. Future research should explore domain-specific augmentation strategies, including atmospheric variation simulation, multispectral band manipulation, and temporal augmentation that incorporates seasonal and diurnal variations. Such approaches could further enhance performance while maintaining semantic consistency specific to remote sensing imagery characteristics.

Addressing the classification bias caused by class imbalance in RSSC tasks, this study investigated the feasibility of using data augmentation to mitigate class imbalance issues. Four architectures, MobileNet-v2, ResNet101, ResNeXt101_32 × 32d, and Transformer, were selected and fine-tuned using VHR imagery from Shinan district and its surrounding areas. A class-imbalanced high-resolution remote sensing image dataset was constructed, and comprehensive data augmentation methods (mirroring, rotation, cropping, HSV perturbation, and gamma transformation) were used to alleviate the class imbalance problem. The impact of class imbalance on classifier performance was systematically analyzed across all architectures. The results demonstrate that data augmentation represents a realistic and effective approach to mitigating class imbalance problems. Classification bias for minority classes was significantly reduced, with overall performance improvements observed across all classifiers. Among the evaluated models, the Vision Transformer exhibited superior robustness to class imbalance, while ResNet101 demonstrated the most consistent performance among CNN architectures.

Future studies should address several identified limitations. First, inter-class similarity between defined semantic categories artificially increases classification difficulty. Subsequent research should optimize category definitions and sample selection criteria to minimize inter-class confusion while maintaining operational relevance. Second, this study primarily compared performance between severely imbalanced and artificially balanced datasets without exploring intermediate imbalance scenarios. Future investigations should systematically examine varying imbalance ratios, including extreme cases with IR exceeding 100 and long-tailed distributions commonly encountered in large-scale remote sensing applications. In addition, the integration of domain-specific augmentation techniques and multilabel classification frameworks represents promising directions for enhancing model robustness while accommodating the inherent semantic complexity of remote sensing scenes.