Predicting Seabed Sand Content Using Machine Learning and Geostatistics

18 machine learning and geostatistical methods were compared for spatial interpolation of seabed sand content. The study, using samples from August 2010, found RFIDS and RFOK to be the most accurate methods across three regions, reducing prediction error by up to 7%. It provides guidelines for improving spatial interpolations of marine environmental data across the Australian Exclusive Economic Zone (AEEZ).

Use Cases

• Benchmarking spatial interpolation methods (e.g., RFIDS, RFOK) for marine sediment prediction.
• Calibrating geostatistical models using secondary environmental variables.
• Producing seabed sand content maps for the Australian Exclusive Economic Zone (AEEZ).
• Improving prediction accuracy through model averaging techniques.

Strengths

• Compares 18 distinct predictive methods, providing a broad methodological benchmark.
• Identifies specific methods (RFIDS, RFOK) that reduced prediction error by up to 7%.
• Based on samples from Geoscience Australia's authoritative Marine Samples Database.

Limitations

• Core dataset metadata (row count, column names, file size) is completely unknown.
• The 'last updated' dates (2026-06-04, 2026-04-04) are implausible future dates, casting doubt on metadata accuracy.
• Primary data is from a single month (August 2010), limiting temporal analysis.

Provenance

Source

Geoscience Australia's Marine Samples Database, via the Australian Ocean Data Network.

Collection Method

Samples extracted and analyzed using machine learning (Random Forest variants) and geostatistical (kriging) interpolation methods.

Time Range

August 2010

Freshness

Metadata lists last updated dates in 2026, which are implausible; the underlying study data is from 2010.

Geography

Australian Exclusive Economic Zone (AEEZ), specifically three study regions within it.