NEWS.md

Version 2.4.0

Changed

• Rebuilt per-cell graphs with flat node arrays and directly constructed compressed sparse row (CSR) edges, retaining integer Dijkstra. Per-thread buffers eliminate per-cell allocations. Together with the caching and BERI improvements below, these changes are about 4x faster for connectedness and PARC and 3x faster for BERI than v2.3 on the bundled example (8 threads); speedups against v2.2 are about twice as large. CPU/Rayon remains the only execution path.
• Cached window area fractions and outer-edge flags in per-call templates; cells clipped by tile boundaries are still computed directly. Edge distances use per-call tables wherever values provably match the per-edge calculation bit for bit, with direct calculation retained for other edges, including skewed or inconsistent level transforms and distances near an f32 rounding boundary. Distances and integer costs are unchanged.
• Batched BERI similarity calculations across all scenarios of each cell, preserving similarity values.
• Fixed each node's outgoing edge order: same-level queen neighbours in the existing order, then fringe links in get_edge_neighbours order. Shortest-path costs are unchanged, but integer-cost ties can select different parents and adjusted path sums, and indicator accumulation follows a different destination order. On the bundled example 2-16% of cells change, by up to about 1e-4 (2.4e-4 relative); grids where most edges round to zero integer cost can change more. Repeated runs and different thread counts remain bit-identical on the same build and platform.
• Improved Rust connectivity() validation: missing transforms are checked before processing, and skipped powers of two produce a clear panic message instead of an index error. Both previously surfaced when a focal cell first reached the affected level. Python connectedness() and beri() already validate both.
• Moved the Rust numerical engine (overviews, window templates, graph building, routing, indicator metrics and pixel coverage) into a separate crate, connectivity-core 0.1.0 in crates/connectivity-core, with a documented array-based Rust API and no Python, GDAL or file I/O dependencies. The rust_conn extension is now a thin PyO3 wrapper around it: Python function signatures, outputs, exceptions and threading are unchanged. The core is versioned and tagged separately (connectivity-core-vX.Y.Z); only vX.Y.Z tags publish the Python package to PyPI.

Version 2.3.0

Changed

• Reduced routing overhead by building one adjacency list for both adjusted and intact searches, caching raw adjusted-distance sums along the selected paths, and avoiding repeated path back-tracing and attribute copying. Integer cost scaling and shortest-path tie handling are unchanged.
• Replaced the remaining standard (SipHash) hash maps in per-cell graph building with FxHashMap, and replaced pathfinding::dijkstra_all with an in-crate Dijkstra that follows the same steps, preserving parents, ties, summation order and outputs. pathfinding is now only a test dependency. Together with the routing optimizations above, these changes made v2.3 about 2x faster than v2.2 in testing.

Fixed

• Made indicator summation deterministic by retaining the adjusted search visitation order instead of iterating over a randomized hash map. Repeated runs on the same build and platform now use a fixed summation order without sorting targets for every focal cell. Final scores may differ in the last few bits from earlier versions because floating-point addition depends on order.
• Integer path costs are now checked for overflow. Inputs that push a path cost past the u32 range (for example a very large max_cost or unscaled resistance values) now raise an error instead of silently wrapping and corrupting the result.

Version 2.2.0

Added

• Added an optional resistance_file argument (and resistance_scale) to connectedness() and beri(), backed by a new traversal argument on the Rust connectivity() binding. When supplied, the resistance raster is decoupled from condition and used only for path traversal — the edge weight becomes w = (1.0 - max_cost) * (1 - resistance) + max_cost — while condition still drives every indicator value (habitat area and the value multiplier). When omitted, condition is used for traversal as before, so output is unchanged. Resistance values are expected in [0, 1] (scaled via resistance_scale); cells valid in condition but missing resistance fall back to condition for traversal and emit a warning, so the analysis domain is never changed silently.
• Added an optional pa_to_pa argument to connectedness() (PARC edition, requires pa_file). When enabled, only graph nodes that fall on a protected area contribute to the connectedness numerator: paths still route through the intervening non-PA landscape (adjusted/intact distances unchanged) and the denominator keeps the full reachable area exactly as when disabled, so only the numerator changes — non-PA destinations contribute nothing and coarse cells straddling a PA boundary contribute only their protected fraction. A protected area that can reach no other scores 0. Defaults to False (original PARC behaviour).
• Added an optional filename argument to pixel_coverage(); when set, the per-pixel coverage array is also written to a GeoTIFF that inherits raster_file's grid and CRS, producing a raster directly usable as the pa_file input to connectedness().
• Added bundled example datasets and an example_data_path() helper (exported from the top-level package) to resolve them by name — site_condition, pa_proportion, and the transgrids/* climate scenarios — so the example notebooks and quick tests run without external downloads.
• Added an example notebook, examples/01_tasmania_indicators.ipynb, demonstrating the connectivity and BERI indicators on the bundled Tasmania data (shipped in the source distribution).

Changed

• Raised the minimum supported Python to 3.10 (requires-python = ">=3.10").
• Updated packaging ([tool.maturin]) to ship the example rasters in both wheels and source distributions, and the example notebooks in the source distribution.
• Published the package on PyPI as eco-connectivity (pip install eco-connectivity), since the name connectivity is taken there; it is still imported as connectivity. Prebuilt wheels for Linux (x86_64, aarch64), macOS (Intel and Apple Silicon) and Windows (x64) mean installing no longer requires Rust. If you installed an earlier version from a locally built wheel, run pip uninstall connectivity first.

Version 2.1.0

Added

• Added make_tile() and make_tiles() helpers for creating non-overlapping raster-pixel core tiles whose internal row/column breaks align to a requested aggregation level, e.g. align_to=max(levels), for tiled connectedness() and beri() runs. The helpers also support optional deterministic cost-balanced planning with balanced=True.

Fixed

• Fixed adjusted path distances to use the same max_cost-scaled edge distances used by Dijkstra routing, avoiding a second hard-coded inverse-condition transform equivalent to max_cost = 2.
• Fixed geographic area weights in tiled runs to use tile-local rows with the already shifted read-window transform, avoiding a double row-offset that could leave residual horizontal tile seams.

Version 2.0.0

Added

• Added a window_mode option to connectedness(), beri(), and the Rust connectivity() binding.
• Added window_mode="circular" for source-centered circular annuli with fractional area/count support at annulus boundaries.
• Added window_mode="square" for source-centered square annuli.
• Added resolution_info() for previewing internally generated level dimensions, resolutions, cell counts, approximate native/km search reach, and small-grid warnings before choosing levels.

Changed

• Changed the default window mode to window_mode="circular".
• connectedness() and beri() now validate that levels form a continuous power-of-two sequence before entering Rust graph construction.
• Updated Rust Python bindings to pyo3 0.28 and numpy 0.28, removing the deprecated gil-refs and py-clone feature usage and migrating to the current PyO3/numpy APIs.
• Updated the Python build configuration for current PyO3/maturin extension-module handling.
• overview_info() is now a backward-compatible wrapper for resolution_info().
• Reduced routing/window-construction overhead by avoiding per-node adjacency vector clones during Dijkstra, hoisting max-level lookup out of per-cell window construction, using faster hash maps for internal graph adjacency data, and writing parallel row results directly into the output array.
• Reduced Python-side overhead by reading BERI scenario rasters concurrently, avoiding an extra temporary array when converting masked raster data to NaN-filled float32 arrays, and applying the option=3 square-root transform in place.

Removed

• Removed window_mode="block" and the original snapped multi-resolution window construction from the public API and Rust implementation.
• Removed automatic grid-effect filtering and the filter_kwargs arguments from connectedness() and beri().
• Removed remove_grid_bias() from the package exports;
• Removed the obsolete Rust inpaint_nans_diffusion Python binding and implementation, which were only used by the previous FFT notch-filter path.

Version 1.1.1

Fixed

• Fixed a PARC-connectedness crash when the protected-area mask leaves no valid cells in the analysis window. Empty windows now return all-NaN output instead of failing during filtering/inpainting.
• Filtered bounded outputs from connectedness() and beri() are now clipped back to the theoretical [0, 1] range after FFT notch filtering. This applies to connectedness (option = 1), connected habitat (option = 2), geometric-mean connected habitat (option = 3), PARC-connectedness, and BERI whenever filtering is enabled.
• Float32 outputs written by write_raster() now set nodata = NaN in the GeoTIFF metadata, and masked float arrays are written with NaN cells so the raster profile matches the stored data.

Version 1.1.0

Added

• Added remove_grid_effect() as a standalone FFT notch-filter post-processing function (also exported from the top-level package import).
• Added a Rust inpaint_nans_diffusion implementation (Rayon-parallel), with optional thread control via n_threads.

Fixed

• Fixed seam artifacts caused by swapped row/column arguments in graph distance coordinate conversion (Affine::xy now consistently receives (row, col)).
• Fixed overview alignment across independently processed tiles by anchoring aggregation windows to global dataset pixel coordinates via tile offsets.
• Fixed cross-level edge-neighbour mapping to be offset-aware (global-anchor consistent) when linking a level to its higher level.
• Fixed occasional missing edge rows/cols in buffered tile reads by adding a half-pixel epsilon to the buffer extent.

Changed

• Graph construction now passes tile offsets through the full path (core -> builder -> graph::fringe) to keep multi-resolution linking globally aligned.
• Node IDs are now tile-invariant (NodeId = u64 from level/global-row/global-col), removing tile-local ID drift between adjacent runs.
• Core traversal now processes overview levels in sorted order.
• Non-closed tile reads are snapped to the coarsest overview grid, improving consistency of aggregated counts/weights at tile boundaries.
• Added Area resampling for 2D overviews (latitude-weighted by cos(phi)), alongside Average, Sum, and Count.
• 3D overview generation now explicitly supports only Average and Sum resampling.
• overview_info() now reports estimated overview dimensions and does not require pre-existing file overviews.
• connectedness() and beri() now apply FFT notch filtering (remove_grid_effect) instead of the previous Gaussian smoother.
• Filtering in connectedness() and beri() is now controlled via filter_kwargs: pass None to disable filtering, or {} to apply default filtering (notch_width=3).
• Inpainting used by filtering now relies on the Rust implementation and uses n_threads when provided.

Removed

• Removed legacy smoothing_filter() backward-compatibility wrapper.
• Removed SciPy-based inpainting fallback and dropped the scipy Python dependency.

Version 1.0.0

Added

• Incorporated habitat area into connectivity calculations using per-node cell counts (aggregate-based weighting).
• Integrated overview generation directly in Rust to eliminate preprocessing steps and reduce raster I/O bottlenecks.
• Added the anyhow crate for improved, centralized error handling.
• Refactored core functionality into a dedicated module for improved structure and maintainability.

Removed

• Removed create_overviews() and has_overview(); overview handling is now fully internal.

Version 0.6.1

• Fixed a bug for reading files with different overview levels.
• Added additional early checks for grid shapes and values.

Version 0.6.0

• Added closed_border argument to control whether cells outside the mask are excluded from the analysis.
• Fixed the polygon-masked output to not return the buffered array.
• Added compilation CPU optimisation flags (slower complie but faster run).
• Changed the overview_levels argument of the create_overviews() function to levels for consistency. Also, file_path in read_raster() function.

Version 0.5.0

• Added has_overview() function for checking overviews.

Version 0.4.2

• A few internal structural changes; e.g., changes to the NumPy array memory layout and added structure in the Rust code for performance improvement.

Version 0.4.1

• Fixed a bug involving a mismatch between conditions and transgrids for BERI by handling NANs.
• Replaced Arc<T> with Rc<T> reference counting, since atomic operations and thread sharing are not needed.
• Improved code readability by adding more structs.

Version 0.4.0

• Added PARC-connectedness calculation to the connectedness() function.
• Fixed a bug in reading transgrids with a polygon mask.

Version 0.3.0

• Fixed the code to accept raster integer data type.
• Added a scaling factor (scale arg) to convert condition data to the 0–1 range.
• The output array/map is now cropped to the supplied polygon mask.
• Added type hints to the main functions.
• Made several improvements on the Rust side, such as using Arc<T> for cheap cloning and introducing the Graph struct for more organised code.

Version 0.2.3

• Fixed isolated pixel calculation using their own values for a synthetic neighbour.
• Increased the capacity of the graph to take more neighbours for high-resolution rasters.
• Changed the default value of the sigma parameter to 1.0.

Version 0.2.2

• Removed the dependency on GDAL, replacing it with rasterio.
• Updated logic to ignore overview level 1 (both when reading and generating) since it corresponds to the original resolution and requires no additional computation.
• Fixed the read_raster() and create_overviews() functions to ensure that the correct overview layers are created and read properly.

Version 0.2.1

• Added Euclidean and Haversine distances for projected and unprojected grids, respectively, instead of using cell units.
• Some changes in the arguments names including max_cost, window_size and outer_window.
• Some small fixes for graph calculation.

Version 0.2.0

Added geographic distance (in km) instead of cell units Replaced the scale argument with max_cost Various minor improvements

Version 0.1.0

• Initial version.