Orthoseg

Improvements

- Add support to train subject on different pixel sizes (143)
- Several small improvements (128)

Bugs fixed

- Fix check that traindataset has validation samples should not be per file (131)
- Fix support for WMS username/password (146)

Deprecations and compatibility notes

- Change default nb_concurrent_calls when downloading layer to 1 instead of 6 (149)

Improvements

- Add support to train and detect on local image file layer (111)
- Add check that train and validation data is mandatory (118)
- Support geopandas 0.13 + remove dependency on pygeos (121)
- Improve performance of prepare_traindatasets for large label files (116)
- Use [pygeoops](https://github.com/pygeoops/pygeoops) for inline simplification during
predict (123):
- 10x faster for inline simplification
- new LANG+ algorithm for improved simplification (= now default algorithm)
- Several small improvements to logging, linting,... (110, 112, 122,...)

Bugs fixed

- Fix occasional "database locked" error while predicting + improve predict performance (120)

Improvements

- Add support for username/password logon for WMS (97)
- Improve performance of load_images (99)

Bugs fixed

- Fix download of model in load_sampleprojects (100)

Improvements

- Avoid warning blockysize is not supported for jpeg when downloading images (77)
- Add __version__ attribute to orthoseg (80)
- Add option to specify reclassify to neighbour query as postprocessing (86, 88)
- Reuse sample_project for tests so it is tested as well (81)
- Improve test coverage (79, 82, 83)
- Add support for shapely 2 (92)

Bugs fixed

- Fix occasional errors when predicting with topologic simplify (69)
- Fix non blocking errors in prediction being ignored/not reported + format error email in html (72, 76)
- Fix error in prediction on ubuntu 22.04 (73)
- Fix error when prediction output dir doesn't exist yet (75)
- Fix reclassify_to_neighbours giving undetermined result when 2+ neighbours are reclassification candidates (84)

Deprecations and compatibility notes

- Remove old model architectures: standard+unet and ternaus+unet (82)

Improvements

- Add option to use topologic simplify (22)
- Add option to specify the minimum probability in the result (23)
- Add option to reclassify (e.g. small) features to the class of its neighbour (30, 36)
- Add check if location BBOXs are of the right size (38)
- Improve handling of empty features in training data (16, 19)
- In the polygons label file, change default class column name to "classname" (39)
- Add traindata type todo (37)
- Add support for training data from layers in different projections (56, 60)
- Add list of invalid geometries + coordinates in train mail (40)
- Add option to ignore wms url in capabilities (20)
- Add support for crs's with switched axes order (50)
- Add nb_parallel parameter to predict (15)
- Add pre-simplify with RDP when using (inline) LANG simplification (29)
- Add option to disable ssl verification when downloading sample projects (64)
- Apply black formatting to comply with pep8 (27)
- Enable running CI tests using github actions (42, 67)
- Support newer versions of used packages (59, 61, 62)

Deprecations and compatibility notes

- Disable default simplify in postprocess (32)
- Command 'scriptrunner' renamed to 'osscriptrunner' (59)

Highlights for this release are a feature to be able to combine bands from different WMS layers to one (3 band) input layer, performance improvements,...

Improvements

- Add feature to combine bands from different WMS layers in one input image layer to segment
- Add option to save model after default 100 epochs even if min. accuracy not reached
- Add option to first train some epochs with the encode layers frozen
- Improve input data checks in prepare_traindatasets
- Introduce "scriptrunner", to have more flexibility when running orthoseg tasks combined with other scripts
- Improve/update documentation + sample project
- Improve prediction performance (5)
- Improve logging
- Update dependencies, eg. tensorflow 2.8, geofileops 0.3, geopandas 0.10,...

Bugs fixed

- When downloading a greyscale layer from a WMS, it was created incorrectly