The accurate detection of small ships based on images or vision is critical for many scenarios, like maritime surveillance, port security, and navigation safety. However, achieving accurate detection for small ships is a challenge for cost-efficiency models; while the models could meet this requirement, they have unacceptable computation costs for real-time surveillance. We propose YOLO-LPSS, a novel model designed to significantly improve small ship detection accuracy with low computation cost. The characteristics of YOLO-LPSS are as follows: (1) Strengthening the backbone’s ability to extract and emphasize features relevant to small ship objects, particularly in semantic-rich layers. (2) A sophisticated, learnable method for up-sampling processes is employed, taking into account both deep image information and semantic information. (3) Introducing a post-processing mechanism in the final output of the resampling process to restore the missing local region features in the high-resolution feature map and capture the global-dependence features. The experimental results show that YOLO-LPSS outperforms the known YOLOv8 nano baseline and other works, and the number of parameters increases by only 0.33 M compared to the original YOLOv8n while achieving 0.796 and 0.831 AP^50:95 in classes consisting mainly of small ship targets (the bounding box of the target area is less than 5% of the image resolution), which is 3–5% higher than the vanilla model and recent SOTA models.

Accurate short-term container throughput volume (CTV) forecasting is crucial for optimizing yard management and port operational efficiency. However, conventional forecasting frameworks typically assume deterministic vessel-side container volumes, neglecting the lag and uncertainty inherent to Electronic Data Interchange (EDI) pre-allocation data in terminal operations. To address these critical gaps, we develop a short-term forecasting framework structured around fine-grained Route-Terminal combination entities, grounded in the real-world operational requirements of South China Container Terminals (SCCTs) in Shenzhen. To enhance predictive robustness against operational volatility, we propose an improved ensemble methodology “Per-item and Global Greedy Ensemble” strategy, which apply greedy ensemble algorithm in different entities and provides a fallback global ensemble for few-record entities. We conduct a comprehensive evaluation of various models, including statistical and state-space models, gradient boosting models, classic deep learning architectures, and time series language models (TSLMs) across the following dual granularities: micro-level entity scale and macro-level daily business aggregation. Experimental results reveal distinct model advantages based on container types as follows: for continuous laden container flows, the proposed per-item greedy ensemble shows competitive performance at the daily aggregation level, whereas tree-based models perform better for sparse and irregular container flows. Overall, the decomposed entity framework and per-item ensemble strategy provide terminal operators with actionable decision support to improve short-term scheduling flexibility and resource utilization in stochastic port environments.