paths:
    out_folder: Directory to store all output results (models, logs, predictions).
    out_model_name: Name identifier for the saved model.
    train_csv: CSV files containing paths and labels for training datasets.
    val_csv: CSV files containing paths and labels for validation datasets.
    test_csv: CSV files containing paths and labels for test datasets.
    global_mtd_folder: Path to folder with global metadata required for processing.
    ckpt_model_path: Path to a pretrained model checkpoint to initialize weights.
tasks:
    train: Enable or disable training phase.
    train_tasks:
        init_weights_only_from_ckpt: Use checkpoint weights for initialization only (no training resume).
        resume_training_from_ckpt: Resume training from the provided checkpoint.
    predict: Run model inference on test set.
    write_files: Save predictions and outputs to disk.
    georeferencing_output: Apply georeferencing to output files.
    metrics_only: Compute and report metrics without running training or prediction.
hyperparams:
    num_epochs: Total number of training epochs.
    batch_size: Number of samples per training batch.
    seed: Random seed for reproducibility.
    learning_rate: Initial learning rate for optimizer.
    optimizer: Choice of optimizer (adamw, adam, or sgd).
    optim_weight_decay: Weight decay regularization.
    optim_betas: Beta coefficients for Adam-based optimizers.
    scheduler: Learning rate scheduler strategy.
    warmup_fraction: Warm-up fraction for one_cycle_lr.
    plateau_patience: Patience for reduce_on_plateau scheduler.
hardware:
    accelerator: Type of hardware to use (gpu, cpu).
    num_nodes: Number of nodes for distributed training.
    gpus_per_node: GPUs used per node.
    strategy: Distributed training strategy (e.g., auto, ddp).
    num_workers: Number of data loading workers.
saving:
    ckpt_save_also_last: Save the final model checkpoint at the end of training.
    ckpt_weights_only: Save only the model weights (not optimizer or scheduler).
    ckpt_monitor: Metric used to determine the best checkpoint.
    ckpt_monitor_mode: Direction to optimize the monitored metric (max or min).
    ckpt_earlystopping_patience: Epochs to wait for improvement before stopping.
    cp_csv_and_conf_to_output: Copy configuration and CSV files to output directory.
    enable_progress_bar: Show training progress bar.
    progress_rate: Update frequency of progress bar (in steps).
    ckpt_verbose: Print detailed checkpoint saving information.
    verbose_config: Print full config details to console.

modalities:
    inputs:
        AERIAL_RGBI       : Enable/disable AERIAL_RGBI modality. [True/False]
        AERIAL-RLT_PAN    : [True/False]
        DEM_ELEV          : [True/False]
        SPOT_RGBI         : [True/False]
        SENTINEL2_TS      : [True/False].
        SENTINEL1-ASC_TS  : [True/False]
        SENTINEL1-DESC_TS : [True/False]
        
    inputs_channels:
        AERIAL_RGBI       : Selected channels for AERIAL_RGBI input. Starts at 1.
        SPOT_RGBI         : Selected channels for SPOT_RGBI input. Starts at 1.
        SENTINEL2_TS      : Selected channels for Sentinel-2 time series. Starts at 1.
        SENTINEL1-ASC_TS  : Selected channels for Sentinel-1 ascending time series. Starts at 1.
        SENTINEL1-DESC_TS : Selected channels for Sentinel-1 descending time series. Starts at 1.
    aux_loss: 
        AERIAL_RGBI       : Apply auxiliary loss to AERIAL_RGBI input if mutliple modalities. [True/False]
        AERIAL-RLT_PAN    : [True/False]
        DEM_ELEV          : [True/False]
        SPOT_RGBI         : [True/False]
        SENTINEL2_TS      : [True/False]
        SENTINEL1-ASC_TS  : [True/False]
        SENTINEL1-DESC_TS : [True/False]
    aux_loss_weight: Multiplier for auxiliary loss before combining with main loss. Default 1.
    modality_dropout: Dropout probability per modality (0 = keep all, 1 = drop all systematically).
        AERIAL_RGBI       : 0
        AERIAL-RLT_PAN    : 0
    pre_processings: 
        filter_sentinel2: Enable filtering of Sentinel-2 based on masks.
        filter_sentinel2_max_cloud : Max acceptable cloud cover in Sentinel-2 [%]. Default 1.
        filter_sentinel2_max_snow : Max acceptable snow cover in Sentinel-2 [%]. Default 1.
        filter_sentinel2_max_frac_cover : Max acceptable fractional invalid coverage in Sentinel-2. Default 0.05.
        temporal_average_sentinel2 : Temporal averaging method for Sentinel-2 (False/monthly/semi-monthly).
        temporal_average_sentinel1 : Temporal averaging method for Sentinel-1 (False/monthly/semi-monthly).
        calc_elevation : Enable calculation oF DTM-DSM.
        calc_elevation_stack_dsm : Stack DSM if elevation calculated.
        use_augmentation: Enable data augmentation for training. Default False.
    normalization: 
        norm_type : Normalization strategy (custom, scaling, without).
        AERIAL_RGBI_means : Mean values used to normalize AERIAL_RGBI input if custom.
        AERIAL_RGBI_stds  : Std. dev. values used to normalize AERIAL_RGBI input if custom.
        AERIAL-RLT_PAN_means : Mean values for AERIAL-RLT_PAN input if custom.
        AERIAL-RLT_PAN_stds  : Std. dev. values for AERIAL-RLT_PAN input if custom.
        SPOT_RGBI_means : Mean values for SPOT_RGBI input if custom.
        SPOT_RGBI_stds  : Std. dev. values for SPOT_RGBI input if custom.
        DEM_ELEV_means : Mean elevation values (DSM, DTM) if custom.
        DEM_ELEV_stds  : Std. dev. of elevation values (DSM, DTM) if custom.

models:
    monotemp_model:  # Encoder-decoder architecture for single-date (monotemporal) inputs
        arch: Encoder-decoder architecture from SMP. Eg, swin_tiny_patch4_window7_224-upernet
        new_channels_init_mode: Strategy to initialize new input channels if above 3 channels (options: copy_first, copy_second, copy_third, random).
    multitemp_model:  # Architecture for handling multi-temporal (time series) inputs
        ref_date: Reference date (MM-DD) for temporal position encoding.
        encoder_widths: Channel sizes at each stage in the encoder; must match decoder.
        decoder_widths: Channel sizes at each stage in the decoder; must match encoder.
        out_conv: Output convolutional layer configuration [intermediate_channels, num_classes].
        str_conv_k: Kernel size for structured convolutions.
        str_conv_s: Stride for structured convolutions.
        str_conv_p: Padding for structured convolutions.
        agg_mode: Aggregation method for temporal features (e.g., att_group).
        encoder_norm: Normalization type used in the encoder (e.g., group).
        n_head: Number of attention heads in the model.
        d_model: Dimensionality of the model features.
        d_k: Key dimensionality for attention mechanism.
        pad_value: Value used for padding missing temporal data.
        padding_mode: Type of padding used in convolutions (e.g., reflect).

labels: # List of tasks
  - AERIAL_LABEL-COSIA
  - ALL_LABEL-LPIS 
labels_configs:
    AERIAL_LABEL-COSIA:  
        task_weight: Weight assigned to this task during multi-task training. Default 1.
        value_name: Mapping of class indices to semantic category names.
            0  : 'building'
            1  : 'greenhouse'
            2  : 'swimming_pool'
            3  : 'impervious surface'
            4  : 'pervious surface'
            5  : 'bare soil'
            6  : 'water'
            7  : 'snow'
            8  : 'herbaceous vegetation'
            9  : 'agricultural land'
            10 : 'plowed land'
            11 : 'vineyard'
            12 : 'deciduous'
            13 : 'coniferous'
            14 : 'brushwood'
            15 : 'clear cut'
            16 : 'ligneous'
            17 : 'mixed'
            18 : 'undefined'
        value_weights:
            default: Default weight for all classes (used in loss calculation). Default 1.
            default_exceptions: Specific class indices assigned 0 weight (ignored).
                15: 0
                16: 0
                17: 0
                18: 0
            per_modality_exceptions: Placeholder for setting modality-specific class weights.
                AERIAL_RGBI:
                  18: 0
    ALL_LABEL-LPIS:  
        task_weight: Weight assigned to this task during multi-task training.
        label_channel_nomenclature: Index of the label channel used.
        value_name: Mapping of class indices to crop type names.
        value_weights:
            default: Default class weight for the entire label set.
            default_exceptions: No exception weights defined (can be added).
            per_modality_exceptions: Placeholder for modality-specific class weight overrides.

# ======================
# I/O
# ======================
output_path: Path to store output results.
output_name: Identifier name for the output files.
write_dataframe: Whether to write slicing geometries of input area. [True/False]
output_type: Format of the output predictions (argmax or class_prob).
cog_conversion: Convert outputs to Cloud-Optimized GeoTIFF format. [True/False]
# ======================
# Model & Inference
# ======================
model_weights: Path to pretrained model weights for inference. [.safetensors / .ckpt / .pth]
use_gpu: Whether to use GPU for inference.
batch_size: Number of samples per inference batch.
num_worker: Number of data loading workers during inference.
img_pixels_detection: Size of input image chunks in pixels.
margin: Overlap margin (in pixels) between tiles.
output_px_meters: Output resolution in meters per pixel.
# ======================
# Model Framework
# ======================
monotemp_arch: SMP Architecture used for monotemporal inference. Eg., swin_base_patch4_window12_384-upernet
multitemp_model_ref_date: Reference date (MM-DD) for aligning multi-temporal inputs.
# ======================
# Modalities (inputs used for inference)
# ======================
modalities:
  inputs:
      AERIAL_RGBI       : Enable AERIAL_RGBI input. [True / False]
      AERIAL-RLT_PAN    : [True / False]
      DEM_ELEV          : [True / False]
      SPOT_RGBI         : [True / False]
      SENTINEL2_TS      : [True / False]
      SENTINEL1-ASC_TS  : [True / False]
      SENTINEL1-DESC_TS : [True / False]
  AERIAL_RGBI:
      input_img_path: Path to AERIAL_RGBI input image.
      channels: Channels used from the input image. Starts at 1.
      normalization:
        type: Normalization method (e.g., custom).
        means: Per-channel mean for normalization.
        stds: Per-channel standard deviation for normalization.
  AERIAL-RLT_PAN:
      input_img_path: Path to AERIAL-RLT_PAN input.
      channels: Channels used.
      normalization:
        type: Normalization method.
        means: Mean value.
        stds: Standard deviation.
  DEM_ELEV:
      input_img_path: Path to DEM input.
      channels: Channels used.
      normalization:
        type: Normalization method.
        means: Per-channel mean.
        stds: Per-channel standard deviation.
      calc_elevation: Whether to compute elevation from DSM/DTM.
      calc_elevation_stack_dsm: Stack derived DSM with the input.
  SPOT_RGBI:
      input_img_path: Path to SPOT imagery input.
      channels: Channels used.
      normalization:
        type: Normalization method.
        means: Per-channel mean.
        stds: Per-channel standard deviation.
  SENTINEL2_TS:
      input_img_path: Path to Sentinel-2 time series input.
      channels: Channels used.
      dates_txt: Path to text file listing image dates.
      filter_clouds: Enable cloud filtering.
      filter_clouds_img_path: Path to Sentinel-2 cloud mask.
      temporal_average: Whether to average the time series over time.
  SENTINEL1-ASC_TS:
      input_img_path: Path to Sentinel-1 ascending orbit input.
      channels: Channels used.
      dates_txt: Path to date list (optional).
      temporal_average: Whether to average over time.
  SENTINEL1-DESC_TS:
      input_img_path: Path to Sentinel-1 descending orbit input.
      channels: Channels used.
      dates_txt: Path to date list (optional).
      temporal_average: Whether to average over time.
# ======================
# Tasks
# ======================
tasks:
  - name: AERIAL_LABEL-COSIA  # Semantic segmentation label task (land cover classes)
    active: Whether to activate this label set during inference. [True / False]
    class_names: Mapping of class indices to category names.
      0: building
      1: greenhouse
      2: swimming_pool
      3: impervious surface
      4: pervious surface
      5: bare soil
      6: water
      7: snow
      8: herbaceous vegetation
      9: agricultural land
      10: plowed land
      11: vineyard
      12: deciduous
      13: coniferous
      14: brushwood
      15: clear cut
      16: ligneous
      17: mixed
      18: undefined
  - name: ALL_LABEL-LPIS  # Crop type classification label task
    active: Whether to activate this label set during inference. [True / False]
    class_names: Mapping of class indices to crop categories.

flair_LC = 
0    : 'building'                'bâtiment'                   '#db0e9a'
1    : 'greenhouse'              'serre'                      '#9999ff'
2    : 'swimming_pool'           'piscine'                    '#3de6eb' 
3    : 'impervious surface'      'zone imperméable'           '#f80c00' 
4    : 'pervious surface'        'zone perméable'             '#938e7b' 
5    : 'bare soil'               'sol nu'                     '#a97101' 
6    : 'water'                   'eau'                        '#1553ae' 
7    : 'snow'                    'neige'                      '#ffffff' 
8    : 'herbaceous vegetation'   'surface herbacée'           '#55ff00' 
9    : 'agricultural land'       'culture'                    '#fff30d' 
10   : 'plowed land'             'terre labourée'             '#e4df7c' 
11   : 'vineyard'                'vigne'                      '#660082' 
12   : 'deciduous'               'feuillu'                    '#46e483' 
13   : 'coniferous'              'conifère'                   '#194a26' 
14   : 'brushwood'               'broussaille'                '#f3a60d' 
15   : 'clear cut'               'coupe'                      '#8ab3a0' 
16   : 'ligneous'                'ligneux'                    '#c5dc42' 
17   : 'mixed'                   'mixte'                      '#6b714f'
18   : 'undefined'               'indéterminé'                '#000000'
flair_LPIS = 
0    : 'Grasses'                 'Herbe prédominante'         '#92d050' 
1    : 'Wheat'                   'Blé'                        '#d7e600' 
2    : 'Barley'                  'Orge'                       '#e0e000' 
3    : 'Maize'                   'Maïs'                       '#fff100'  
4    : 'Other cereals'           'Autres céréales'            '#ffff00' 
5    : 'Rice'                    'Riz'                        '#e8e8e8'
6    : 'Hemp/Flax/Tobacco'       'Lin/Chanvre/Tabac'          '#dceaf7' 
7    : 'Sunflower'               'Tournesol'                  '#d29ead' 
8    : 'Rapeseed'                'Colza'                      '#d29ed0' 
9    : 'Other oilseed crops'     'Autres oléagineux'          '#ffbe99' 
10   : 'Soy'                     'Soja'                       '#ffc000' 
11   : 'Other protein crops'     'Autres protéagineux'        '#ff9000' 
12   : 'Fodder legumes'          'Légumineuse fourragères'    '#009999' 
13   : 'Beetroots'               'Betterave'                  '#808000' 
14   : 'Potatoes'                'Pomme de terre'             '#a7a700' 
15   : 'Other arable crops'      'Autres arables'             '#89896d' 
16   : 'Vineyard'                'Vigne'                      '#f2cfee' 
17   : 'Olive groves'            'Oliveraie'                  '#6f6633' 
18   : 'Fruit orchards'          'Verger fruits'              '#ac8141' 
19   : 'Nut orchards'            'Verger noix'                '#996633' 
20   : 'Other permanent crops'   'Autres pérenne'             '#80c1d7' 
21   : 'Mixed crops'             'Mélange'                    '#000000' 
22   : 'Background'              'Autres'                     '#000000'