Configuration

DeepTrees uses Hydra for configuration. Users must a configuration file called config.yaml in the root folder of the project. We will continue to add more parameters in the future as we parameterize the code further with Hydra.

Below you can find the default configuration for training and inference. You can use this as a template for your own configuration file.

  • Inference default configuration (object): Cannot contain additional properties.

    • name (string, required): Identifier for the run. Will be used in folder names and logs. Default: “inference-halle”.

    • output_dir (string, required): Output folder. Default: “results”.

    • pretrained_model (string, required): Path to pretrained model(s). You can pass a single model or a list of pretrained models. In the latter case, their predictions will be averaged. Default: “/work/ka1176/caroline/gitlab/deeptrees/experiments/finetune-halle/2024-11-25_09-32-21/unet-halle_jitted.pt”.

    • logdir (string, required): Set by hydra. This is where we find logs and results. Default: “${hydra.run.dir}”.

    • model_name (string, required): Short and memorable model name. Default: “unet-halle”.

    • seed (null, required): Random seed. If null, it is not fixed. Default: null.

    • crs (string, required): Coordinate reference system. Default: “EPSG:25832”.

    • polygon_file (string, required): Output file to store the segmented polygons. Default: “treecrowns.sqlite”.

    • baumkataster_file (string, required): Path to the file containing the Baumkataster ground truth. Default: “/work/ka1176/shared_data/2024-ufz-deeptree/halle-baumkataster/itc005211130105323_point.shp”.

    • baumkataster_intersection_file (string, required): Output file to store the polygons that intersect with Baumkataster. Default: “treecrowns_baumkataster.sqlite”.

    • callbacks: Refer to Callbacks.

    • data: Refer to Data.

    • model: Refer to Model.

    • trainer: Refer to Trainer.

    • hydra: Refer to Hydra.

  • Callbacks (object): Callbacks to pass to the trainer during inference. Cannot contain additional properties.

  • Data (object): Cannot contain additional properties.

    • _target_ (string, required): class to instantiate. Default: “deeptrees.dataloading.datamodule.TreeCrownDelineationDataModule”.

    • rasters (string, required): Path to folder containing the raster tiles. Default: “/work/ka1176/shared_data/2024-ufz-deeptree/polygon-labelling/pool_tiles”.

    • masks (string, required): Only for consistency. Default: null.

    • outlines (string, required): Only for consistency. Default: null.

    • distance_transforms (string, required): Only for consistency. Default: null.

    • batch_size (integer, required): Test batch size (must be 1). Default: 1.

    • test_indices (array, required): Optional list of test indices. If given, only these rasters are used for prediction. Default: null.

    • divide_by (integer, required): Value by which to divide the input rasters. Default: 255.

    • dilate_outlines (integer, required): Number of pixels by which to dilate the outlines. Default: 0.

    • num_workers (integer, required): Number of workers in the Pytorch DataLoader. Default: 8.

    • ndvi_config: Refer to NdviConfig.

    • augment_eval: Refer to AugmentEval.

    • ground_truth_config: Refer to GroundTruthConfig.

  • AugmentEval (object): Cannot contain additional properties.

    • Pad: Refer to Pad.

  • Pad (object): Cannot contain additional properties.

    • padding (integer, required): Padding to apply to all sides of the input raster. This is currently hard-coded 500-> 512! Default: 6.

  • GroundTruthConfig (object): Cannot contain additional properties.

    • labels (string, required): Only for consistency. Default: null.

  • NdviConfig (object): Cannot contain additional properties.

    • concatenate (boolean, required): Concatenate NDVI to RGBI. Default: true.

    • rescale (boolean, required): Rescale NDVI to [0, 1]. Default: false.

    • red (integer, required): Index of red channel in raster. Default: 0.

    • nir (integer, required): Index of infrared channel in raster. Default: 3.

  • Model (object): Cannot contain additional properties.

    • _target_ (string, required): Class to instantiate. Default: “deeptrees.model.deeptrees_model.DeepTreesModel”.

    • num_backbones (integer, required): Number of models to average. This will be overwritten if pretrained_model is a list. Default: 1.

    • in_channels (integer, required): Number of input channels (e.g. RGBI+NDVI). Default: 5.

    • architecture (string, required): TreeCrownDelineation architecture. Default: “Unet”.

    • backbone (string, required): TreeCrownDelineation backbone. Default: “resnet18”.

    • apply_sigmoid (boolean, required): If True, apply sigmoid to mask and outline outputs to return probability maps. Default: false.

    • postprocessing_config: Refer to PostprocessingConfig.

  • PostprocessingConfig (object): Cannot contain additional properties.

    • min_dist (integer, required): Minimum distance between neighbouring tree crowns. Default: 10.

    • mask_exp (integer, required): Parameter for feature extraction. Default: 2.

    • outline_multiplier (integer, required): Parameter for feature extraction. Default: 5.

    • outline_exp (integer, required): Parameter for feature extraction. Default: 1.

    • dist_exp (number, required): Parameter for feature extraction. Default: 0.5.

    • area_min (integer, required): Minimum area for a polygon to be considered. Default: 3.

    • sigma (integer, required): Gaussian filter standard deviation in feature extraction. Default: 2.

    • label_threshold (number, required): Minimum height of local maxima during feature extraction. Default: 0.5.

    • binary_threshold (number, required): Threshold value for the feature map, lower is background. Default: 0.1.

    • simplify (number, required): Polygon simplification distance, vertices closer than this value are simplified. Default: 0.3.

    • active_learning (boolean, required): Calculate mean entropy per tile. Default: true.

    • save_entropy_maps (boolean, required): Save the entropy heatmaps. Default: true.

    • save_predictions (boolean, required): Save the predictions (mask, outline, distance transform). Default: true.

  • Trainer (object): Cannot contain additional properties.

    • _target_ (string, required)

    • devices (integer, required): Number of GPUs to use in parallel. Default: 1.

    • accelerator (string, required): Choose GPU if available. Default: “auto”.

    • enable_progress_bar (boolean, required): Enable progress bar. Default: true.