Champion Chasing Boy Team’s Solution for the 2020 Digital China Innovation Contest – Digital Government Track

Competition Background

The 2020 Digital China Innovation Contest (DCIC 2020) focused on digital government, encouraging AI and big‑data applications to improve maritime governance in Fujian Province. The challenge required participants to identify fishing vessel operation types (trawl, surrounding net, or stabbing net) from Beidou positioning data, with additional AIS data in later rounds.

Problem Statement & Data

Participants received 11,000 vessel trajectory records (7,000 training, 2,000 testA, 2,000 testB) containing anonymized vessel IDs, coordinates (x, y), speed, heading, timestamp, and a label for the operation type. AIS data with unique device IDs were also provided for cross‑modal matching.

Feature Engineering

Both single‑attribute and multi‑attribute features were constructed. Single‑attribute features captured speed, heading, and coordinate statistics at global and local levels using quantiles and bucketed aggregates. Multi‑attribute features focused on cross‑features of speed and direction.

Algorithm Framework

The core model was LightGBM, trained separately for vessels with stable versus highly variable trajectories. Additional matching algorithms aligned Beidou and AIS tracks using Time‑Weighted Similarity (TWS) and Space‑Weighted Similarity (SWS). An adversarial verification approach transformed the unsupervised matching problem into a supervised one, allowing integration of multiple information sources.

System & Application Architecture

The visualization front‑end was built with Angular.js and ECharts, enabling a clear separation of concerns and secure data handling. The back‑end used Flask served through Nginx, employing pre‑computation, caching, and indexing to ensure responsive performance.

Visualization & Knowledge Graph

Interactive dashboards display vessel heatmaps, real‑time positioning, and safety risk scores. A maritime knowledge graph aggregates vessel profiles, operational habits, and equipment status, supporting government monitoring and collision‑avoidance alerts.

Safety Monitoring & Prediction

Time‑series models forecast vessel locations for the next 30 minutes, feeding into a collision‑risk index that considers speed, density, signal loss, and positional drift. Heatmaps of predicted traffic enable proactive safety warnings and resource allocation.

Conclusion

The presented solution combines AI modeling, multi‑source trajectory matching, robust feature engineering, and end‑to‑end visualization to accurately classify fishing vessel activities and provide actionable safety insights for maritime governance.