Wildfires are among the most devastating natural disasters, causing significant damage to the environment, the economy, wildlife, and human lives, and requiring an early detection system. This paper presents a novel LightFire-Net deep cross model fusion architecture for Wildfire classification and interpretation from remote sensing images. The proposed model is computationally efficient and employs Efficient Flame Pattern blocks with depthwise separable convolutional layers, a Lightweight spectral spatial module, and Efficient Feature Fusion for real-time wildfire classification and detection. The model's multi-stage architecture enables hierarchical feature extraction, resulting in more accurate and efficient classification. Bayesian Optimization (BO) is employed to select hyperparameters, thereby improving the training of the proposed model on selected remote sensing datasets. A Softmax classification layer is employed for the final prediction. The proposed model achieves exceptional performance with above 98% prediction accuracy while maintaining only 1.37 million parameters and an inference speed of 7.74 ms. A comprehensive evaluation across multiple dataset splits (90:05:05-40:30:30) demonstrates robust generalization, with accuracy ranging from 98.02% to 99.25%. The five-fold mean accuracy further confirms the model's stability, at 98.62%±0.15%. The proposed model outperforms existing CNN and transformer-based architectures while being significantly more efficient, making it highly suitable and reliable for real-time monitoring in resource-constrained environments.
Aerial scene recognition has progressed substantially with deep learning methods for RGB and hyperspectral imagery; however, existing approaches typically operate on single modalities or rely on explicit multimodal fusion, limiting scalability, flexibility, and deployment in heterogeneous sensing environments. To address this limitation, we propose a sensor-agnostic semantic representation learning framework that formulates multimodal learning as the unification of semantic representations rather than feature-level fusion. The proposed architecture employs modality-specific encoders and projection heads to map spatial and spectral–spatial features into a shared semantic embedding space, enabling modality-invariant representation learning while preserving discriminative characteristics of each sensing modality. A composite objective integrating cross-spectral alignment, intra-class compactness regularization, and prototype-based semantic anchoring is introduced to enforce consistent embedding geometry and improve class separability across modalities. A unified classifier operating within this shared space enables reliable inference from a single modality input without requiring paired data or explicit fusion. Extensive evaluations on multiple benchmark datasets, including Houston 2013 for cross-modality RGB–hyperspectral analysis, UC Merced for independent RGB aerial scene classification, and Indian Pines for hyperspectral land-cover recognition, demonstrate the robustness and generalization capability of the proposed framework. In Houston 2013, the method achieves 96.4% (RGB) and 97.3% (hyperspectral) overall accuracy, with cross-modality transfer performance of 87.2% (RGB → HSI) and 88.7% (HSI → RGB), further improving to 97.0% and 97.8% under joint training. On UC Merced and Indian Pines, the model attains 98.7% and 97.6% overall accuracy, respectively. These results establish semantic representation unification as a scalable and effective alternative to conventional multimodal fusion for heterogeneous remote sensing environments.
1 January 2026 in IEEE Transactions on Consumer Electronics
Consumer vehicles and aftermarket devices embed cameras, inertial sensors and GNSS, enabling road-condition awareness as a consumer electronics service. Yet manual surveys and isolated sensing remain costly and inconsistent at city scale, while road traffic injuries impose a socio-economic burden often estimated at about 3% of GDP. Single-source approaches fail under field conditions because camera monitoring degrades in low illumination and occlusion, inertial sensing is sensitive to vehicle dynamics and speed variation and satellite positioning can be sparse or unstable for precise defect localisation. This paper proposes a unified vehicle-mounted sensing and edge computing pipeline that integrates visual, inertial and positioning streams and couples reliability-aware alignment and fusion with communication-efficient reporting for decision making. The system learns geo temporal representations through self-supervised pretraining, synchronises asynchronous streams online using reliability-weighted Dynamic Time Warping and performs hierarchical transformer fusion with reliability gating to control cross-sensor dependencies under time-varying quality. Window-level predictions are converted into compact georeferenced events that can be transmitted over bandwidth-limited links and aggregated into Geographic Information System ready heatmaps and priority segments. On the Passive Vehicular Sensors dataset, the proposed approach achieves 0.91 accuracy and 0.90 macro averaged F1, improving over early fusion at 0.84 accuracy. The results indicate that reliability-aware multimodal learning improves robustness and enables consumer-grade road monitoring with actionable geospatial outputs.
15 June 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Aerial scene classification, which aims to automatically label an aerial image with a specific semantic category, is a fundamental problem for understanding high-resolution remote sensing imagery. For accurate classification of aerial scenes from remote sensing images, image classification must effectively integrate local spatial details, context-dependent information from global regions, and hierarchical representations of different scenes. This study proposes a hybrid multitemporal information fusion architecture for aerial scene recognition from Remote Sensing images. The proposed model combines the modified version of EfficientNet-B0 (truncated to a maximum number of convolutional layers), Vision Transformer (ViT), and FractalNet to leverage the strengths of these three models. Each model's branch runs in parallel and is fused via depth-wise feature concatenation. Extensive experiments are conducted on four benchmark datasets (EuroSAT, NWPU, EarthHazards, and Turkey Earthquake), demonstrating the integrated model's improved performance over traditional CNN and transformer models, as well as its compact nature (approximately 2.8 million parameters) compared to these models. Extensive ablation studies also validate the effectiveness of the combined approaches of block-depth selection, self-attention layer, and fusion strategy. The results from the combined model show that it is an effective and efficient approach to classify large numbers of Remote Sensing viewpoint scene types.