Jiangbing Sun

Rock image recognition based on deep learning serves as a critical task in intelligent geological analysis and core documentation. Although recent advances in deep learning have established the technical foundation for rock image identification, its practical deployment remains constrained by two main bottlenecks: intricate feature representations and ubiquitous label noise. Complexity of rock image features limits the effectiveness of intelligent model applications. Additionally, manual mislabeling and coarse marking issues during lithology annotation are often overlooked, leading to label noise in training datasets. To resolve these challenges, this study proposes a robust learning framework that integrates a novelty-driven feature extraction module with noise-resistant loss functions. Specifically, a Rock Hierarchical Heterogeneous Attention (RHHA) module is designed for enhancing rock feature extraction capabilities by applying distinct spatial and channel attention biases at shallow and deep layers of the network. Furthermore, Symmetric Cross Entropy (SCE) loss is used for loss calculation to enhance the robustness of models under noisy conditions. Comparative experiments were carried out on a real-world dataset of metamorphic rock images from northern Jiangsu Province, China. The results show that the proposed training framework achieves the best performance in rock image identification tasks, outperforming baseline and other comparative methods. The RHHA module outperforms other existing attention modules in capturing the texture and semantic features of rocks. In particular, the SCE loss effectively mitigates overfitting and maintains excellent generalization capability in the presence of noisy labels. The proposed framework holds promise for providing new insights into rock image recognition tasks.

Soil erosion poses a severe threat to agricultural productivity and ecological security on the Loess Plateau. However, previous studies have rarely integrated physical modeling, elasticity coefficients, and spillover effects into a unified framework at the county level. To address this gap, this study coupled the Revised Universal Soil Loss Equation (RUSLE) with the Spatial Durbin Model (SDM) to systematically investigate the spatiotemporal dynamics, factor elasticity characteristics, and spatial dependence mechanisms of soil erosion on the Loess Plateau from 1990 to 2024. Results show that the annual average erosion rate decreased by 15.5%, with a highly volatile phase before 2001 and a stabilized, low-erosion phase thereafter. The driving factors exhibited marked heterogeneity in direction and strength. The land cover and management factor (C) was the strongest erosion-reducing factor, whereas annual precipitation (PRE) was the primary natural erosion-enhancing factor. County-level erosion also displayed significant positive spatial dependence. PRE had a stable positive indirect effect, whereas C and the support practice factor (P) mainly contained erosion within local jurisdictions. These findings of a unified RUSLE–SDM framework reveal a joint driving mechanism of localized human interventions and climate-driven cross-regional spillovers, providing quantitative support for differentiated soil and water conservation strategies on the Loess Plateau.

Aquaculture ponds and salt pans represent the dominant forms of coastal production spaces along the Jiangsu coast, China; however, their long-term co-evolution and mutual transitions remain poorly understood. To bridge this gap, this study developed a 40-year (1985–2025) spatiotemporal dataset of these land covers leveraging Landsat imagery via the Google Earth Engine (GEE) platform. We established an integrated classification workflow encompassing single-scene water mask extraction, annual Modified Normalized Difference Water Index (MNDWI)-based water frequency statistics, Otsu automatic thresholding, connected-component labeling, and the masking of natural water bodies. The resulting dataset demonstrated high reliability, achieving overall accuracies (OA) ranging from 92.32% to 94.15% and an average Kappa coefficient of 0.89. Based on multi-metric analyses of area dynamics, annual change rates, and transition patterns, we identified three distinct co-evolutionary stages: simultaneous expansion (1985–1995), internal reorganization (1995–2015), and overall contraction (2015–2025). Notably, transitions between the two production spaces were highly asymmetric over the 40-year period; the area converted from salt pans to aquaculture ponds was approximately 15.23 times greater than the reverse conversion. Furthermore, their distribution exhibited strong spatial heterogeneity at the county level, underscoring the critical role of localized coastal planning in balancing economic production and wetland conservation. Ultimately, this work provides foundational data and methodological insights for long-term coastal ecological monitoring and sustainable production space management.

The scientific data of the Earth surface system is characterized by multi-source heterogeneity and dynamic correlation, so constructing an efficient data association network and enabling intelligent knowledge services is a hot topic. Nevertheless, confronted with the existing challenges of onerous data acquisition, inadequate precision of data recommendation, excessive time and labor consumption, as well as insufficient semantic reasoning in intelligent question-and-answer (Q&A) systems, we propose an intelligent framework that integrates dynamic optimization and retrieval-augmented generation (RAG) technology to address the problems of strong subjectivity in the setting of edge weight thresholds in association networks and insufficient semantic inference in intelligent Q&A. First, a multidimensional association network is constructed based on metadata features, redundant edge pruning is achieved through dynamic threshold analysis, and key nodes are identified by combining complex network centrality theory to optimize network structure and storage efficiency. Secondly, the RAG-based intelligent Q&A model is designed to transform the association triples into a paragraph-based knowledge base, generate a domain Q&A dataset using a large language model GPT-4o, and fine-tune the word embedding model to improve the semantic representation accuracy. Experiments show that the number of network edges is reduced by about 70% after optimization, and the node importance analysis accurately identifies key data nodes; the fine-tuned model improves each index by 6% on average in the retrieval task, and the Q&A system significantly outperforms the traditional method in terms of indexes such as relevance and completeness. This study provides innovative solutions for the intelligent service of scientific data in Earth surface systems and promotes the deep integration of association networks and generative AI.

Precise object detection is fundamental to robust indoor navigation and localization. However, the practical deployment of deep learning-based detectors on mobile platforms is frequently impeded by their extensive parameter counts, substantial computational overhead, and prolonged inference latency, rendering them impractical for real-time and GPU-independent applications. To overcome these limitations, this paper presents Nav-YOLO, a highly optimized and lightweight architecture derived from YOLOv8n, specifically engineered for navigational tasks. The model’s efficiency stems from several key improvements: a ShuffleNetv2-based backbone significantly reduces model parameters; a Slim-Neck structure incorporating GSConv and GSbottleneck modules streamlines the feature fusion process; the VoV-GSCSP hierarchical network aggregates features with minimal computational cost; and a compact detection head is designed using Hybrid Convolutional Transformer Architecture Search (HyCTAS). Furthermore, the adoption of Inner-IoU as the bounding box regression loss accelerates the convergence of the training process. The model’s efficacy is demonstrated through a purpose-built Android application. Experimental evaluations on the VOC2007 and VOC2012 datasets reveal that Nav-YOLO substantially outperforms the baseline YOLOv8n, achieving mAP50 improvements of 10.3% and 5.0%, respectively, while maintaining a comparable parameter footprint. Consequently, Nav-YOLO demonstrates a superior balance of accuracy, model compactness, and inference speed, presenting a compelling alternative to existing object detection algorithms for mobile systems.