Zhan Yang

Deep-time Digital Earth (DDE) intends to harmonize the deep-time earth data and open geoscience knowledge. To assist the objective, efficient cross-modal retrieval methodologies are essential to process heterogeneous earth observation data. Hash learning maps diverse modalities data into binary codes, providing a viable solution for large-scale remote sensing data retrieval, while the Transformer architecture enhances the accuracy and relevance of generated models by dynamically integrating knowledge in an unsupervised paradigm. In consequence, the combination of Transformer and hash learning replaces the dense vector search with Hamming distance, significantly reducing storage requirements and increasing retrieval efficiency via compact binary codes. In this paper, we propose a novel unsupervised Contrastive Transformer Hashing approach for cross-modal Remote Sensing data retrieval, denoted as CTHRS. Initially, a Transformer-based multi-modal fusion network is developed, which bridges the semantic gap between remote sensing images and texts via self-attention mechanism and KL divergence alignment. Subsequently, a neighborhood structure preserving mechanism is presented, which maintains the intra-modal semantic coherence through dual similarity matrices derived from both original and hashing representations. Furthermore, a contrastive hash learning objective is designed, which combines intra-modal decomposition and cross-modal alignment losses. Comprehensive experiments on RSICD, RSITMD, and UCM datasets demonstrate the superior performance of CTHRS to state-of-the-art hashing approaches. The codes and datasets are available at: https://github.com/holdibson/CTHRS.

Group Emotion Recognition (GER) aims to estimate the collective affect displayed by multiple people in a shared scene. Hard-label voting and static fusion strategies used in prior work often overlook the intrinsic ambiguity of individual emotions and the non-linear group-level interactions, including conflict, tendency and resonance. We present Fuzzy Emotional Hierarchical Shared Space (FEHSS), a three-subspace framework that assigns each person a soft membership vector across emotion categories, simultaneously optimises conflict suppression, tendency regulation and resonance enhancement, and updates its fusion rules on the fly through a hierarchical reinforcement optimiser (HIRO). Comprehensive evaluations on GroupEmoW, GAD 2.0, and EMOTIC-7 confirm that FEHSS sets a new state of the art, reaching 91.77% accuracy on GroupEmoW and consistently outperforming previous methods on the other two benchmarks. Extensive ablation studies confirm the complementary roles of the three subspaces and the necessity of dynamic parameter tuning. Qualitative visualisations further illustrate how FEHSS resolves contradictory cues and preserves subtle shared emotions that rigid fusion schemes overlook. Additional experiments on the EMOTIC-7 dataset with seven emotion classes show that FEHSS attains 74.3% top-1 accuracy and 71.8% macro-F1. The proposed framework thus offers both higher accuracy and stronger interpretability, providing a new pathway toward robust GER in unconstrained real-world scenarios.