International Journal of Advanced AI Applications

A Review of Applications of Deep Learning-Driven Multi-Source Remote Sensing Spatiotemporal Fusion in Wetland Monitoring

2025-12-03T08:09:09+00:00

Wetlands are vital ecosystems on Earth, and the dynamic monitoring of their changes is crucial for ecological conservation, biodiversity maintenance, and water resource management. Multi-source remote sensing spatiotemporal fusion technology can integrate the advantages of datasets with different spatiotemporal resolutions to generate image sequences with high spatiotemporal resolution, providing core data support for dynamic wetland monitoring. By virtue of its powerful feature extraction and nonlinear fitting capabilities, deep learning has greatly improved the accuracy and adaptability of spatiotemporal fusion, driving wetland monitoring toward refinement and dynamization. This paper systematically sorts out the research background and significance of deep learning-driven multi-source remote sensing spatiotemporal fusion technology, categorically elaborates on mainstream fusion models, summarizes their application achievements in four core scenarios (wetland boundary extraction, type classification, ecological parameter inversion, and dynamic change monitoring), analyzes the key problems faced by current technologies, and prospects future research directions, thereby providing a reference for the in-depth application of this technology in wetland monitoring.

A Hybrid TAM–Learning Analytics Framework for Predicting University Students’ Adoption of Educational Technology

2025-11-24T14:07:25+00:00

This study develops a hybrid analytical framework integrating the Technology Acceptance Model (TAM) with learning analytics indicators to explain university students’ adoption of educational technology. While TAM emphasizes perceptual constructs such as perceived usefulness (PU) and perceived ease of use (PEOU), modern digital learning systems generate rich behavioral data that may also shape learners’ adoption decisions. Data were collected from 162 undergraduate students using validated measurement scales for PU, PEOU, and behavioral intention (BI), together with self-reported learning analytics indicators including interaction frequency, time-on-task, and digital participation levels. Structural equation modeling was conducted using Python based SEM analysis. Results show that PU and PEOU significantly predict BI, consistent with classical TAM. Incorporating learning analytics indicators improves explanatory power, with hierarchical regression revealing an increase in R2 from 0.689 (TAM-only) to 0.739 in the hybrid model (∆R2 = 0.050). Random Forest analysis further confirms the predictive importance of PU and learning analytics features. These findings demonstrate that behavioral engagement data substantially enhance students’ technology adoption processes. The study contributes theoretically by integrating cognitive and behavioral perspectives of adoption, and offers practical implications for designing more engaging and data-informed digital learning environments.

Adaptive-Gated Spiking Neural Networks with Memristive Crossbars for Real-Time Athlete Injury Prediction

2025-11-28T14:52:33+00:00

We propose a Spiking Neural Network (SNN) with adaptive gating for real-time athlete damage prediction,addressing the inefficiencies of conventional deep learning in processing multimodal sensor data. Traditional methods often suffer from high computational overhead due to redundant modality processing,whereas our approach dynamically gates irrelevant inputs early in the pipeline,significantly reducing energy consumption without compromising accuracy. The core innovation lies in a spike-based gating mechanism that evaluates contextual relevance of each modality,selectively suppressing low-importance signals through learnable coefficients. Furthermore,early multimodal fusion is achieved via dendritic compartments,enabling event-driven computation at the spike level,which naturally aligns with the sparse and asynchronous nature of sensor data. The architecture is co-designed with memristive neuromorphic hardware,where synaptic weights are mapped to analog conductances,thereby minimizing energy per spike through in-memory computation. Experimental results demonstrate that the proposed system operates at under 5 mW while maintaining competitive prediction performance,making it suitable for wearable deployment. The integration with existing sensor infrastructure is seamless,as the SNN replaces traditional deep learning layers without requiring modifications to preprocessing or decision support modules. This work bridges the gap between neuromorphic computing and practical sports analytics,offering a scalable solution for real-time injury risk assessment. The combination of adaptive gating, hardware-aware optimization, and multimodal fusion establishes a new direction for energy-efficient deep learning in edge applications.