30 July 2024 in IEEE Transactions on Geoscience and Remote Sensing
Recent approaches based on Neural Radiance Fields (NeRF) showcase remarkable results in 3D reconstruction of small-scale scenes by encoding volume density and color observations using implicit functions. However, when confronted with complex and diverse large-scale scenes, it invariably experiences issues such as blurry textures and missing details. In this work, we present a superpixel-based neural radiance field (named SuperNeRF), an additional loss for learning radiance fields that takes advantage of superpixels texture constraints. Building upon NeRF, we leverage superpixels to establish spatial consistency constraints, enabling the precise extraction of 3D geometry and appearance for large-scale scenes. SuperNeRF is capable of guiding locally adjacent and similar pixels to form nearly consistent ray termination distributions, and it is compatible with the state-of-the-art NeRF-based methods. Comprehensive experiments conducted on representative aviation and aerospace datasets demonstrate that our SuperNeRF exhibits a significant superiority in accuracy over state-of-the-art methods. Code will be available at https://github.com/xczbecalm/supernerf.
17 February 2026 in IEEE Transactions on Geoscience and Remote Sensing
Semantic urban 3D meshes have widely applied in the fields of photogrammetry and remote sensing. Acquiring discriminative feature representations to improve semantic segmentation accuracy in urban 3D meshes remains a significant challenge. To address this issue, a novel framework for urban 3D mesh semantic segmentation is proposed in this paper, which comprises three key innovations: (1) an adaptive local edge feature extractor designed to capture fine-grained geometric details across diverse local structures; (2) a local-global fusion module that synergistically combines these local features with global contextual representations extracted via transformer networks; and (3) a supervised category-center-aware contrastive learning strategy that addresses class imbalance and improves the discriminability of feature representations. Ablation studies validate the individual contributions of each component. Evaluated on three benchmark datasets, namely SUM, seMet, and Wuhan, our framework achieves state-of-the-art performance, consistently surpassing existing methods across all three datasets. The code is available at https://github.com/NJTechZhang/Category-Center-Aware-Contrastive-Learning-Mesh.git.
27 March 2026 in IEEE Geoscience and Remote Sensing Letters
Self-supervised learning (SSL) has shown strong potential for learning representations from unlabeled data, yet its direct application to remote sensing (RS) imagery remains challenging due to complex backgrounds and severe target-scale variations. To address these challenges, we propose Context-Enhanced Semantic Contrast (CESC), a self-supervised pre-training framework tailored for RS imagery. CESC integrates masked reconstruction, explicit contextual enhancement, and cross-scale semantic contrast to jointly model fine-grained spatial structures and scale-invariant semantic representations. Extensive experiments on scene classification and semantic segmentation benchmarks demonstrate that CESC consistently outperforms state-of-the-art SSL methods under equivalent settings, validating its effectiveness and transferability for RS representation learning.