Lonboard

Lonboard follows an architecture where all TypeScript models are combined into a single entry point. This entry point is compiled by ESBuild into an ES Module.

• Loading: The Python Map class loads this JavaScript bundle via the _esm key, which informs Jupyter/ipywidgets where to find the bundle.
• Synchronization: anywidget and ipywidgets handle the serialization of data from Python to JavaScript, ensuring both sides stay in sync automatically.

Use ipyleaflet if you need a broad range of data types and fine-grained control over rendering for small-to-medium datasets.

Lonboard is preferred for large datasets because ipyleaflet relies on GeoJSON for data transfer, which is slow to write, read, and large in transit. Lonboard avoids these bottlenecks by using more efficient data transport.

Use kepler.gl-jupyter if you want a high-level, browser-based visualization and analysis toolkit, or if you need to create standalone static HTML files containing datasets.

Lonboard is preferred if you perform most analysis in Python and require maximum rendering performance. Lonboard lacks the browser-based UI for data exploration found in kepler.gl but offers better performance because it avoids the large GeoJSON-like JavaScript objects and text serialization used by kepler.gl.

Pyodide environments have stricter memory limits than standard Python environments. To prevent memory exhaustion when using Lonboard in the browser, explicitly delete Python objects that are no longer needed using the del statement.

# Example of manual memory management in Pyodide
del large_data_object

Lonboard's performance and functionality are built upon four foundational technologies:

• deck.gl: A JavaScript geospatial data visualization library that uses the GPU for high-performance rendering of large datasets.
• GeoArrow: A memory format for efficient, uncompressed representation of geospatial vector data.
• GeoParquet: A file format for efficient encoding and decoding of geospatial vector data using compression.
• anywidget: A framework used to build custom Jupyter widgets.

Lonboard uses GeoPandas as the primary user interface, managing the internal conversion from GeoPandas to GeoArrow and GeoParquet to facilitate data transport to the browser.

The render_tile parameter in RasterLayer.from_geotiff acts as a custom, client-side tile server running in your Python environment. As you pan or zoom the map, Lonboard automatically calls this callback to fetch and process tiles.

Callback Requirements:

• Input: An async_geotiff.Tile object.
• Output: An instance of lonboard.raster.EncodedImage.
• Capabilities: You can use any Python library (like NumPy or PIL) within the callback to manipulate pixel data, apply band math, or use machine learning models before returning the encoded image.

Supported Formats: While you can return various formats, common choices are PNG, JPEG, or WebP.

Lonboard implements data-driven styling by serializing computed data rather than JavaScript functions.

In other libraries (like pydeck), users often pass strings containing JavaScript callbacks (e.g., "@@=properties.value") to define styling. In Lonboard, you use familiar Python objects like Numpy arrays, Pandas Series, or PyArrow Arrays to define accessors.

Workflow:

1. You compute a property in Python (e.g., using a Matplotlib colormap or an ML model output).
2. Lonboard stores this as a binary array/Parquet file.
3. The binary array is sent to the frontend.
4. The frontend uses the binary buffer directly for rendering.

This approach allows you to use the full power of the Python ecosystem (Matplotlib, Scikit-learn, etc.) for styling without learning a JavaScript DSL.