/landcover_scripts

Scripts for landcover project - https://github.com/microsoft/landcover

Primary LanguageJupyter Notebook

Scripts for landcover web-tool project

Experimentation procedure

  1. Obtained 1922 correction labels.

    • Among which 180 were duplicates.
    • Resulting with 1742 correction labels.
    • Labels were distributed as follows:
      1. Chicken house - 576
      2. Built - 387
      3. Field - 338
      4. Tree Canopy - 275
      5. Water - 166

  2. Generate tuned model

    • Command ran:

    python generate_tuned_model_v2.py --in_geo_path ../notebooks/all_corrections_no_dups.geojson --in_model_path ../landcover-old/web_tool/data/naip_autoencoder.h5 --in_tile_path ../landcover-old/web_tool/tiles/m_3807537_ne_18_1_20170611.mrf --out_model_path ./naip_autoencoder_tuned_uneven.h5 --num_classes 2 --gpu 1

    python generate_tuned_model_v2.py --in_geo_path ../notebooks/all_corrections_no_dups.geojson --in_model_path ../landcover-old/web_tool/data/naip_autoencoder.h5 --in_tile_path ../landcover-old/web_tool/tiles/m_3807537_ne_18_1_20170611.mrf --out_model_path ./naip_autoencoder_tuned_even.h5 --num_classes 2 --gpu 1

    python generate_tuned_model_v2.py --in_geo_path ../notebooks/all_corrections_no_dups.geojson --in_model_path ../landcover-old/web_tool/data/naip_demo_model.h5 --in_tile_path ../landcover-old/web_tool/tiles/m_3807537_ne_18_1_20170611.mrf --out_model_path ./naip_demo_tuned_uneven.h5 --num_classes 2 --gpu 1

    python generate_tuned_model_v2.py --in_geo_path ../notebooks/all_corrections_no_dups.geojson --in_model_path ../landcover-old/web_tool/data/naip_demo_model.h5 --in_tile_path ../landcover-old/web_tool/tiles/m_3807537_ne_18_1_20170611.mrf --out_model_path ./naip_demo_tuned_even.h5 --num_classes 2 --gpu 1

  3. Sampling

    • For even sampling: 150 samples each -> 750 total

    • For uneven sampling: 375 Chicken houses, 375 random samples from the other classes.

    Example distribution of uneven sampling:

    4.0 375 3.0 129 2.0 123 1.0 73 0.0 50

    Example distribution of even sampling:

    4.0 150 3.0 150 2.0 150 1.0 150 0.0 150

Main Project Link