Neutrosophic soft-computing ensemble for high-accuracy satellite image scene classification

1 Introduction

Image classification is one of the core computer-vision tasks that categorises images and assigns semantic labels, enabling effective interpretation of visual data. Compared to image segmentation and object detection, scene-level classification offers faster processing and a robust approach, making it suitable for rapidly categorising heterogeneous and complex datasets (Liu et al., 2020). In recent years the field has shifted from traditional Convolutional Neural Networks (CNNs) to more sophisticated architectures, including vision transformers.

Among the vast applications of image classification, satellite image classification remains particularly significant owing to its applicability across agricultural monitoring, disaster management, urban planning, and land-use/land-cover mapping (Thapa et al., 2023). Deep learning has gained prominence over the last few years due to its strong representational capacity, but three structural challenges persist: (i) high inter-class visual similarity, (ii) intra-class spatial variance caused by varying sensor orientations and imaging conditions, and (iii) the computational burden of processing high-resolution aerial tiles (He et al., 2016).

Within the deep learning paradigm, three architectural families have established themselves as the most powerful tools for remote-sensing scene recognition. First, ConvNeXt (Liu et al., 2022) re-examines the design space of pure convolutional networks by incorporating transformer-inspired choices - large 7 × 7 depthwise kernels, GELU activation, and layer normalisation - while retaining the simplicity of a standard CNN. Second, the High-Resolution Network (HRNet) (Wang et al., 2020) maintains parallel multi-resolution feature streams throughout the network, preserving fine spatial detail that standard downsampling architectures inevitably lose. Third, Swin Transformer (Duan et al., 2025) extends the Vision Transformer (ViT) paradigm through hierarchical, shifted-window self-attention that captures long-range contextual dependencies at sub-quadratic computational cost.

Each architectural family encodes a distinct inductive bias: local texture hierarchies (ConvNeXt), spatial precision at multiple resolutions (HRNet), and global scene-level context (Swin). This observation motivates the central hypothesis of this paper: a principled combination of all three architectures, fused through uncertainty-aware soft computing, substantially outperforms any single architecture regardless of scale or tuning.

While the present study focuses on optical RGB scene-level classification, the core principles of architectural complementarity and neutrosophic uncertainty modelling are conceptually applicable to broader remote sensing tasks including SAR-based target detection, multi-scale object screening, and quantity-aware vessel surveillance; these extensions are discussed in Section 2.5.

Soft-computing methods - fuzzy logic (Zadeh, 1965), intuitionistic fuzzy logic (Atanassov, 1986; Krassimir and Parvathi, 1986), and neutrosophic logic (Smarandache, 1998) - provide a theoretically grounded framework for modelling and resolving prediction uncertainty in multi-source decision problems. Neutrosophic logic is particularly attractive because it independently parameterises Truth ( T ), Indeterminacy ( I ), and Falsity ( F ) without constraining their sum, enabling explicit quantification of inter-branch disagreement as an I signal used to down-weight uncertain predictions.

The primary contributions of this work are.

• 1. A triple-branch ensemble integrating ConvNeXt-Small, HRNet-W18, and Swin Transformer (Small) selected for their complementary inductive biases relevant to aerial scene classification.
• 2. A comprehensive comparison of three soft-computing fusion strategies—Normal Fuzzy, Intuitionistic Fuzzy, and Neutrosophic—benchmarked against simple averaging and against all individual backbone models.
• 3. A novel neutrosophic fusion formulation that uses inter-branch probability deviation as an explicit Indeterminacy measure, implementing uncertainty-aware soft voting.
• 4. State-of-the-art results on the UC Merced dataset: 99.29% accuracy and 99.31% precision.

The remainder of the paper is organised as follows. Section 2 reviews related work. Section 3 describes the dataset and preprocessing. Section 4 presents the methodology. Section 5 reports experimental results. Section 6 discusses findings. Section 7 concludes.

2 Related work

2.1 CNN-based remote sensing classification

Convolutional Neural Networks have long served as the backbone for remote-sensing scene classification, providing strong baselines through local texture extraction. The Remote Sensing Transformer (TRS) framework (Zhang et al., 2021) proposed a progressive architectural strategy—from pure CNNs through CNN-Transformer hybrids to pure transformers—demonstrating that integrating transformer modules into CNN backbones substantially boosts performance. ConvNeXt (Liu et al., 2022) challenged the transformer paradigm by modernising ResNet through 7 × 7 depthwise convolutions, GELU activation, and layer normalisation. For remote sensing, STConvNeXt (Liu et al., 2025) reduced the parameter footprint via a Split-based Mobile Convolution module, while combining ConvNeXt-Tiny with Efficient Channel Attention and label smoothing (Perdana and Arimurthy, 2024) demonstrated measurable accuracy gains. A hybrid learnable fusion of ConvNeXt and Swin Transformer (Qezelbash and Hicklin, 2025) further confirmed the complementarity of convolutional and attention-based representations.

2.2 Vision transformer architectures

Vision Transformer (ViT) (Dosovitskiy et al., 2020) redefined image processing by treating images as sequences of non-overlapping patches processed through multi-head self-attention. STMSF (Duan et al., 2025) extended the Swin Transformer (Liu et al., 2021) for remote sensing with a multi-scale feature fusion module and a spatial attention pyramid network to handle the wide variation in object scales inherent in aerial imagery. The Lightweight Dual-Branch Swin Transformer (LDBST) (Zheng et al., 2023) reduced computational cost by pairing a lightweight CNN branch for local features with a ViT branch enhanced by Conv-MLP layers. For hyperspectral data, Spectral Swin Transformer (Liu et al., 2023) adapted shifted-window attention to operate along the spectral dimension, grouping correlated bands to reduce redundancy. The Spectral-Spatial Swin Transformer (SSST) (Peng et al., 2023) coupled CNN-based spatial feature extraction with reinforced spatial position encoding to jointly preserve spectral characteristics and pixel spatial coordinates. The hybrid HDLEM-FFTSRSI model (Oqaibi, 2026) integrated EfficientNetV2-S, ViT, and ConvNeXt for environmental monitoring on remote-sensing imagery.

2.3 High-resolution networks

Standard deep architectures progressively reduce spatial resolution, fundamentally losing spatial precision. HRNet (Wang et al., 2020) addresses this by maintaining parallel branches at different resolutions with repeated multi-scale feature exchange throughout the entire forward pass. SE-HRNet (Li et al., 2020) enhanced HRNet with Squeeze-and-Excitation blocks that model inter-channel dependencies for complex remote-sensing scenes. Lite-HRNet (Yu et al., 2021) replaced standard pointwise convolutions with conditional channel weighting of linear complexity, achieving competitive accuracy for edge deployment. Liu and Zhu (Vandanapu et al., 2025) applied a modified HRNet to bi-temporal aerial imagery for land-cover change detection, exploiting spatial precision to identify fine-scale boundary changes. The Hybrid HRNet-Swin Transformer (Algarni et al., 2025) coupled HRNet for local detail with Swin for long-range dependencies, outperforming state-of-the-art methods on aerial segmentation. MSHRNet (Chen et al., 2025a) introduced a multi-scale high-resolution architecture for land cover classification from high-spatial-resolution imagery.

2.4 Fuzzy logic and soft computing in remote sensing

Zadeh’s seminal paper (Zadeh, 1965) introduced fuzzy sets, replacing binary membership with a continuous grade in [0, 1] to model inherently gradual class boundaries. Klir and Yuan (Klir and Yuan, 1995) provided the definitive treatment of fuzzy set operations and inference mechanisms. Dubois and Prade (Dubois and Prade, 2012) analysed fuzzy measures for imprecise and incomplete information. Bloch (Bloch, 1994) demonstrated that treating pixel intensities as fuzzy sets effectively captures image imprecision, particularly in segmentation and edge detection. Nedeljkovic (Nedeljkovic, 2004) applied fuzzy classification to remote sensing, addressing the mixed-pixel problem by assigning partial membership to multiple classes simultaneously. Seasonal Contrast (SeCo) (Manas et al., 2021) demonstrated that domain-aligned self-supervised pretraining from unlabelled satellite data further improves baseline performance beyond direct ImageNet transfer.

2.5 Broader remote sensing applicability

While the present work focuses on optical RGB scene-level classification, the uncertainty-aware triple-branch ensemble paradigm carries broader implications for remote sensing tasks that involve higher spatial complexity, alternative sensing modalities, and finer-grained recognition objectives.

In high-resolution aerial perception, progressive multi-scale screening paradigms have emerged as effective strategies for transitioning from coarse scene-level labels toward fine-grained object-level decisions. The progressive ship detection paradigm of (Zhang et al., 2025), which follows a divergence-to-concentration and population-to-individual hierarchy, demonstrates that coarse-to-fine architectural organisation substantially improves detection performance in complex SAR scenes. The HRNet branch in the proposed framework, which maintains parallel multi-resolution feature representations throughout the entire forward pass, is architecturally compatible with such hierarchical pipelines: its highest-resolution branch outputs could naturally seed a detection head, while the neutrosophic Indeterminacy signal could act as a gating mechanism to suppress uncertain region proposals in cluttered scenes.

For SAR imagery, ensemble diversity has proven equally critical. The triple-level sparsity awareness framework for marine ship surveillance (Zhang et al., 2026) demonstrates that multi-level feature modelling significantly improves detection robustness in the presence of speckle noise and geometric distortion. However, direct transfer of the proposed framework to SAR modalities is non-trivial: optical-domain ImageNet-pretrained backbones exhibit substantial representational degradation when applied to speckle-corrupted SAR data. This represents a clear limitation of the current work when extended beyond the optical domain, and domain-adapted pretraining or dedicated speckle-aware preprocessing stages would be prerequisites for any such extension.

The neutrosophic logic foundation of the proposed fusion strategy also shares conceptual alignment with uncertainty-aware detection approaches. The focus and super-resolution-guided tiny target detection method of (He et al., 2026), which transitions from fuzzy global representations to clear local features, parallels the neutrosophic Truth-Indeterminacy-Falsity decomposition: both frameworks explicitly address the challenge of low-confidence, ambiguous predictions by separating reliable evidence from uncertain evidence before making a final decision. In zero-shot and small-target scenarios, the Indeterminacy component I b ( c ) defined in Equation 9 is similarly well-positioned to detect and down-weight anomalously spread softmax distributions caused by unknown or unfamiliar target categories.

Finally, extending the ensemble to SAR ocean scene understanding tasks such as panoptic segmentation (Deng et al., 2026) would require augmenting the current classification head with a segmentation decoder. The tri-state prototype self-distillation approach for SAR ocean imagery demonstrates that multi-prototype, multi-state learning strategies can capture the structural complexity of maritime scenes, a challenge that the three-branch complementary design of the proposed ensemble is philosophically aligned with. While outside the scope of the present study, the architectural modularity of the triple-branch framework makes such extensions a natural direction for future work. Together, these directions confirm that the core principles demonstrated here—architectural complementarity and neutrosophic uncertainty modelling—are transferable across remote sensing tasks, though modality-specific adaptation will be required in each case.

3 Dataset and preprocessing

3.1 UC Merced Land Use Dataset

The UC Merced Land Use Dataset comprises 2,100 high-resolution aerial RGB images drawn from the USGS National Map Urban Area Imagery database. The dataset spans 21 semantically distinct land-use categories—agricultural, airplane, baseball diamond, beach, buildings, chaparral, dense residential, forest, freeway, golf course, harbor, intersection, medium residential, mobile home park, overpass, parking lot, river, runway, sparse residential, storage tanks, and tennis court—each containing exactly 100 images at 256 × 256 pixels with a ground sampling distance of approximately 0.3 m per pixel.

Three documented classification challenges make this benchmark rigorous: (i) high inter-class visual similarity—dense residential, medium residential, and mobile home park share overlapping structural and textural patterns; (ii) intra-class spatial variance—building orientations, vegetation density, and shadow patterns vary substantially within a single category; and (iii) complex multi-object compositions—categories such as intersections and overpasses simultaneously contain roads, vegetation, and built structures from multiple semantic domains.

3.2 Data partitioning

A stratified 80/20 train-validation split allocates 1,680 images (80 per class) to training and 420 images (20 per class) to validation. Stratification ensures proportional class representation in both partitions, preventing class-imbalance artefacts from biasing training loss or inflating validation accuracy.

3.3 Preprocessing pipeline

A standardised eight-stage preprocessing pipeline is applied to all images prior to model ingestion, as illustrated in Figure 1. The pipeline is described below.

• 1. Raw Image. The original 256 × 256 RGB aerial patch is used as input.
• 2. Grayscale Conversion. The RGB image is converted to a single-channel grayscale image using standard luminosity weighting.
• 3. Binarisation. Adaptive thresholding is applied to produce a binary image, separating foreground structures from the background.
• 4. Noise Removal. Morphological opening with a 3 × 3 structuring element eliminates isolated noise artefacts without distorting object boundaries.
• 5. Rotation. Images are randomly rotated by up to ± 20 ° with bilinear interpolation to simulate variation in satellite sensor azimuth.
• 6. Cropping. A random crop is extracted at a scale of 0.80–1.0 of the image area to introduce positional variation.
• 7. Centering. The cropped patch is centred and padded to produce a fixed 256 × 256 canvas.
• 8. Resizing. The image is bilinearly resized to 224 × 224 pixels for compatibility with ImageNet-pretrained backbone weights. All images are then channel-normalised using ImageNet mean μ = [ 0.485 , 0.456 , 0.406 ] and standard deviation σ = [ 0.229 , 0.224 , 0.225 ].

Steps (e)–(g) constitute data augmentation and are applied to training images only; validation images receive steps (h) and normalisation only.

4 Proposed methodology

4.1 Triple-branch ensemble architecture

The proposed framework consists of three independently trained deep-learning backbone branches. Each branch is initialised with ImageNet-1K pretrained weights and fine-tuned end-to-end on the UC Merced training split. Each branch produces a 21-dimensional class probability vector via softmax normalisation. These three probability vectors are subsequently fused using one of the soft-computing strategies described in Section 4.2.

ConvNeXt-Small Branch. ConvNeXt (Liu et al., 2022) is derived from ResNet-50 through systematic transformer-inspired updates: (i) a 4 × 4 stride-4 patchify stem; (ii) 7 × 7 depthwise convolutions replacing 3 × 3 convolutions; (iii) Layer Normalisation; (iv) GELU activation; and (v) an inverted bottleneck structure with 4 × channel expansion (≈ 50M parameters). The large depthwise kernels capture repetitive local texture patterns, the dominant discriminative cue in texture-driven UC Merced categories such as agricultural, chaparral, forest, and beach.

HRNet-W18 Branch. HRNet (Wang et al., 2020) maintains parallel feature representations at four spatial resolutions throughout the entire forward pass, beginning with a high-resolution branch at the input resolution and progressively adding lower-resolution branches. All branches exchange information through repeated multi-scale fusion modules at each stage. The W18 configuration uses channel widths of 18, 36, 72, and 144 across the four parallel branches, with the classification output derived from the highest-resolution branch (≈ 29M parameters). This design preserves fine spatial detail decisive for geometry-driven categories: runway, river, harbor, and overpass.

Swin Transformer (Small) Branch. Swin Transformer (Liu et al., 2021) divides the input into non-overlapping 4 × 4 patches and computes multi-head self-attention within local windows of 7 × 7 patches. In alternating layers the window partition is shifted by half the window stride, enabling cross-window interaction without quadratic global attention complexity (≈ 50M parameters). The Swin branch captures global spatial arrangement patterns—layout-defined categories such as golf course, parking lot, and storage tanks—that local convolutional filters cannot model.

All three branches are trained with: AdamW optimiser; cosine annealing learning-rate schedule; initial learning rate 1 × 1 0 − 4; weight decay 1 × 1 0 − 2; cross-entropy loss; batch size 32; maximum 50 epochs with early stopping (patience 10, based on validation accuracy).

4.2 Soft-computing fusion strategies

Let C = { c 1 , … , c 21 } denote the class set. For an input image x, each branch b ∈ { ConvNeXt , HRNet , Swin } produces a softmax probability vector P b ∈ R 21, where P b ( c ) ∈ [ 0,1 ] and ∑ c P b ( c ) = 1. The ensemble prediction iswhere S ( c ) is the fused score under a given strategy.

y ̂ = arg max c ∈ C S c , (1)

Simple Probability Averaging (Baseline).

S avg c = 1 3 P ConvNeXt c + P HRNet c + P Swin c . (2)

Averaging treats all branches equally regardless of inter-branch agreement.

Normal Fuzzy Ensemble. Following Zadeh (1965) and Klir and Yuan (1995), each branch probability P b ( c ) is first mapped through a nonlinear sigmoid-based membership function to reflect the gradual, non-crisp nature of class boundaries in aerial scenes:

μ b c = 1 1 + exp − k P b c − 0.5 , (3)

where k = 10 is a sharpness parameter that amplifies high-confidence predictions toward 1 and suppresses low-confidence predictions toward 0. Unlike direct probability averaging, this transformation is nonlinear: it compresses mid-range probabilities toward 0.5 and sharpens near-certain predictions, introducing a genuine analytical distinction from Simple Probability Averaging. The fused score is then computed as

S fuzzy c = 1 3 ∑ b μ b c . (4)

This formulation treats each softmax output as a fuzzy truth degree and aggregates the outputs after membership shaping, producing a fused score that differs analytically from plain probability averaging.

Intuitionistic Fuzzy Ensemble. Intuitionistic Fuzzy Sets (IFS), introduced by Atanassov (Atanassov, 1986), independently assign to each element with respect to class c: a membership degree μ ( c ), a non-membership degree ν ( c ), and a hesitation margin π ( c ) = 1 − μ ( c ) − ν ( c ) ≥ 0. For each branch b:

μ b c = P b c , ν b c = 1 − P b c , π b c = 0 . (5)

Fusion is performed via the Intuitionistic Fuzzy Weighted Averaging (IFWA) operator (Atanassov, 1999):

μ IFS c = 1 − ∏ b 1 − μ b c 1 / 3 , ν IFS c = ∏ b ν b c 1 / 3 , (6)

S IFS c = μ IFS c − ν IFS c . (7)

Neutrosophic Ensemble (Proposed). Neutrosophic logic (Smarandache, 1998) introduces three fully independent components T , I , F ∈ [ 0,1 ] without the constraint T + I + F = 1. For each class c and branch b.

T b c = P b c , (8)

F b c = 1 − P b c , (9)

I b c = | P b c − 1 3 ∑ b ′ P b ′ c | . (10)

The Indeterminacy component I b ( c ) explicitly quantifies inter-branch disagreement: when all branches agree, I b ( c ) ≈ 0; when branch b diverges, I b ( c ) is large.

The neutrosophic fused score is

S NS c = 1 3 ∑ b T b c ⋅ 1 − I b c . (11)

The ensemble prediction is given in Equation 1. Simple Probability Averaging is defined in Equation 2. The fuzzy membership function and fused score are given in Equations 3, 4. The Intuitionistic Fuzzy components and IFWA operator are defined in Equations 5–7. The Neutrosophic Truth and Falsity mappings are defined in Equations 8, 9, the Indeterminacy component in Equation 10, and the neutrosophic fused score in Equation 11.

Multiplying each branch’s Truth value by ( 1 − I b ( c ) ) down-weights predictions with high inter-branch disagreement and amplifies high-consensus predictions - uncertainty-aware soft voting. Figure 2 illustrates the visual effect of each fusion strategy on a representative aerial image.

5 Experimental results

5.1 Evaluation metrics

All models are evaluated on the 420-image UC Merced validation set across six metrics: overall accuracy (OA), error rate (ER), macro-averaged precision (P), macro-averaged recall (R), macro-averaged F1-score (F1), and average cross-entropy loss ( L ). Macro-averaging computes per-class values and averages them uniformly across all 21 classes, ensuring equal contribution regardless of class sample size.

5.2 Overall performance comparison

Table 1 presents the complete performance comparison across all eleven evaluated configurations: five individual backbone models, three ensemble baselines using simple averaging, and three soft-computing fusion configurations.

TABLE 1. Classification performance of individual models and ensemble configurations on the UC Merced Land Use Dataset validation set (420 images, 21 classes). Metrics: OA = overall accuracy; ER = error rate; P = macro precision; R = macro recall; F1 = macro F1-score; L = average cross-entropy loss. All values are expressed in %; L is dimensionless. The best value in each column is highlighted in bold; † denotes the proposed method.

5.3 Analysis of individual backbone models

Among the five individual backbone models, ConvNeXt-Tiny achieves the highest standalone accuracy of 96.90%, followed by ConvNeXt-Small (96.43%). This advantage validates the effectiveness of large-kernel depthwise convolutions and layer normalisation for local texture discrimination (Liu et al., 2022). HRNet-W18’s relatively lower individual accuracy (92.62%) reflects its narrower channel configuration in the W18 variant; crucially, it contributes distinct representational information to the ensemble—spatial precision rather than texture—which is the diversity property required for effective combination.

5.4 Ablation study

Table 2 presents a systematic ablation study isolating the contribution of each architectural branch and fusion strategy to overall classification accuracy. The stepwise progression confirms two additive sources of improvement: (i) an architectural diversity gain of + 1.67 percentage points from the three-branch combination over the best individual model, and (ii) an uncertainty-modelling gain of up to + 0.72 percentage points from soft-computing fusion over simple averaging.

TABLE 2. Ablation study showing the incremental contribution of each backbone branch and fusion strategy to overall accuracy on the UC Merced validation set. C = ConvNeXt-Small; H = HRNet-W18; S = Swin Transformer (Small); F = soft-computing fusion applied. Δ = gain relative to the best individual model (ConvNeXt-Tiny, 96.90%).

Bold values indicate the highest overall vaccuracy (OA) in each configuration group.

5.5 Comparison with prior methods on the UC merced dataset

Table 3 contextualises the proposed method against representative prior works on the UC Merced dataset. The proposed Neutrosophic Ensemble achieves the highest accuracy among the methods compared here, surpassing the nearest competitor by 0.54 percentage points. It should be noted that two prior methods in Table 3 - Ensemble Hybrid Model and CNN-ViT Ensemble Fusion - do not disclose their train/validation split or evaluation protocol. On a dataset of 2,100 images, differences in split ratio or the use of cross-validation versus a fixed split can produce accuracy variations of 0.3–0.8 percentage points; direct numerical comparison with these methods should therefore be interpreted with appropriate caution.

TABLE 3. Comparison with representative prior methods on the UC Merced Land Use Dataset. Train/val split, pretraining strategy, and evaluation protocol are reported where available to enable fair comparison. NR = Not reported by original authors. Best accuracy shown in bold.

5.6 Computational complexity analysis

Table 4 compares the parameter count and floating-point operations (FLOPs) of each individual backbone against the proposed triple-branch Neutrosophic Ensemble at 224 × 224 input resolution.

TABLE 4. Computational complexity of individual backbone models and the proposed Neutrosophic Ensemble. FLOPs are computed at 224 × 224 input resolution. † Total ensemble cost assumes sequential single-GPU inference across three branches.

Bold values indicate the lowest parameter count and FLOPs among the individual backbone models.

The triple-branch ensemble incurs a total parameter cost of 129.1M, representing a 2.6 × increase over the largest individual backbone (ConvNeXt-Small, 50.2 M) and approximately 3 × the FLOPs, providing a reliable proxy for relative inference cost. This overhead is the direct consequence of maintaining three independently parameterised branches and constitutes the primary practical limitation for latency-sensitive deployment scenarios. The neutrosophic fusion layer itself adds negligible cost over simple averaging, as it consists solely of element-wise operations on the three probability vectors. While exact throughput depends on hardware configuration, on multi-GPU hardware the three branches can be evaluated in parallel, recovering throughput proportionally. As noted in Section 6, knowledge distillation from the ensemble into a single lightweight student network (Yu et al., 2021) represents the most direct path toward preserving accuracy while reducing inference cost for edge deployment.

6 Discussion

Architectural Complementarity. The performance gains in Table 1 are directly attributable to the distinct inductive biases of the three selected architectures. ConvNeXt-Small’s large depthwise kernels capture repetitive local texture patterns decisive in texture-dominant categories (agricultural, chaparral, forest, beach). HRNet-W18’s multi-resolution parallel branches preserve fine spatial details decisive for geometry-dominant categories (runway, river, harbor, overpass). Swin Transformer’s shifted-window attention captures global spatial arrangement, enabling classification of layout-dominant categories (golf course, parking lot, storage tanks). The ablation study confirms that the three-branch combination outperforms all two-branch combinations, indicating non-redundant contributions from each branch.

Neutrosophic Advantage. The Neutrosophic Ensemble’s superiority in precision (99.31% vs. 99.22% for Normal Fuzzy) is attributable to the Indeterminacy component I b ( c ) defined in Equation 10. This component provides a data-driven measure of inter-branch disagreement that down-weights uncertain predictions proportionally to their divergence from the ensemble consensus. For operationally critical remote-sensing applications—land-use change detection, disaster zone mapping, agricultural parcel classification—minimising false positives is frequently more important than maximising recall, making the Neutrosophic Ensemble the preferred configuration.

7 Limitations

Four limitations warrant acknowledgement. First, the empirical evaluation is conducted exclusively on the UC Merced dataset, which contains only 100 images per class and is considered near-saturated in the current deep learning era. With a stratified 80/20 split, the validation partition contains just 20 images per class, meaning the 95% confidence interval on the reported 99.29% overall accuracy is approximately ±1.0 percentage point (Wilson score interval), and accuracy differences smaller than 0.7 percentage points between methods are not statistically distinguishable from this set alone. High accuracy on a small, fixed split may partially reflect benchmark-specific behaviour rather than genuine representational generalisation; cross-validation over multiple random splits would yield more statistically robust estimates. Validation on larger benchmarks such as AID (Xia et al., 2017) (10,000 images, 30 classes) and NWPU-RESISC45 (Cheng et al., 2017) (31,500 images, 45 classes) is therefore necessary before broader claims of generalisation can be made. Second, all backbones were pretrained on natural images; domain-adapted pretraining via SeCo (Manas et al., 2021) may further improve baseline performance. Third, the neutrosophic Indeterminacy formula is hand-crafted; learning fusion weights end-to-end via a meta-learning network is a promising future direction. Finally, maintaining three independent ≈ 50M backbone models increases inference cost; knowledge distillation into a single lightweight student model (Yu et al., 2021) could preserve accuracy gains while enabling edge deployment.

8 Conclusion and future scope

This paper presented a triple-branch deep ensemble framework for satellite image scene classification combining ConvNeXt-Small, HRNet-W18, and Swin Transformer (Small) to jointly exploit local texture hierarchies, high-resolution spatial precision, and global self-attention context. An eight-stage preprocessing pipeline was developed to standardise aerial image inputs. Three soft-computing fusion strategies—Normal Fuzzy logic, Intuitionistic Fuzzy logic, and Neutrosophic logic—were evaluated against a simple averaging baseline and all individual backbone models on the UC Merced Land Use Dataset (21 classes, 2,100 images).

The proposed Neutrosophic Ensemble achieved 99.29% accuracy and 99.31% precision on the UC Merced benchmark, outperforming all individual models and simpler ensemble configurations evaluated in this study. Three principal conclusions emerge. First, architectural complementarity is a high-value resource: combining local texture (ConvNeXt), spatial precision (HRNet), and global context (Swin) yields a + 1.67 percentage point accuracy gain over the best individual model. Second, soft-computing fusion adds + 0.72 percentage points over simple averaging by exploiting inter-branch uncertainty signals. Third, neutrosophic Truth-Indeterminacy-Falsity decomposition achieves the highest precision. These conclusions are drawn from a single benchmark dataset; generalisation to larger datasets and alternative remote sensing modalities remains to be confirmed in future work.

Future work will focus on: (i) validating on AID and NWPU-RESISC45; (ii) incorporating SeCo domain-adapted pretraining; (iii) learning neutrosophic fusion weights through a meta-learning network; (iv) knowledge distillation into a single lightweight deployment model; and (v) extension to multispectral and hyperspectral imagery.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Tianwen Zhang, Southwest Jiaotong University, China

Reviewed by: Yucong He, Southwest Jiaotong University, China

Dunyun He, Southwest Jiaotong University, China