Exploring Spatiotemporal Data Management with Cassandra, GeoMesa, and GeoTrellis

With the widespread adoption of 5G, spatiotemporal data such as delivery logs, location traces, and remote‑sensing information has become a critical data type for many applications.

The talk focuses on the challenges of managing such data in traditional NoSQL databases like Cassandra and introduces a series of techniques and tools to address them.

Spatiotemporal Basics – Spatiotemporal data can be classified into vector data, raster data, point‑cloud data, and trajectory data. Approximately 80% of modern data has a spatial or temporal component, making it essential for data fusion and predictive analytics.

Database Spatiotemporal Engine – A middleware layer that provides unified interfaces for spatial queries, consisting of four core elements: spatial objects, spatiotemporal indexes, serialization, and query execution.

Space‑Filling Curves (SFC) – To fit Cassandra’s key‑value model, multidimensional spatiotemporal queries are transformed into one‑dimensional key ranges using curves such as Z‑curve and Hilbert curve, enabling efficient range scans.

GeoHash Encoding – Latitude and longitude are recursively bisected to produce a binary representation that is then converted to an integer hash, allowing hierarchical precision control and fast proximity searches.

GeoMesa – An open‑source library built on NoSQL stores for vector data management and large‑scale analytics. It provides ETL tools, Kafka integration, and supports Spark and MapReduce. GeoMesa stores data in four tables (attr, id, z2, z3) to separate attributes, identifiers, spatial, and spatiotemporal indexes. To mitigate data skew, keys include a random byte, a time‑slice, the GeoHash, and a feature ID.

Cassandra Storage Scheme – Example tables for the GDELT dataset illustrate how events with latitude, longitude, and timestamp are stored across attr, id, z2, and z3 tables. Queries first locate the appropriate shard, then scan the relevant Z‑range, and finally retrieve the feature object using the feature ID.

GeoMesa Spark Integration – GeometryRDD converts spatial features into Spark RDDs, enabling distributed processing and analysis of geospatial data.

GeoTrellis – Focuses on raster data. It provides a conversion framework that leverages Spark to transform raster tiles into RDDs, which can then be stored or further processed. The workflow includes reprojection, tiling, space‑filling curve encoding, and persistence.

RasterFrame – A platform that combines Python and Spark for remote‑sensing image analysis, offering geospatial extraction capabilities.

Ganos – Alibaba Cloud’s native spatiotemporal engine that supports various underlying stores and provides a unified API for managing spatiotemporal datasets.

The presentation concludes with a summary of the demonstrated pipelines and invites the audience to join the Cassandra community for further discussion.