U-Net-for-Tithe-Map-LULC

All the files needed to train an improved U-Net model for vectorising land cover based on boundary line information in Tithe Maps for historic land use and cover mapping

Making training data:

• trace boundaries in QGIS and save polyline vector layer as a geopackage
• run \training-data\CreateMatchinglMasksAndImage.ipynb to create matching/overlapping images for the georeferenced raster map scan and the traced features
• run \training-data\PatchifyTitheMap.ipynb to patchify the tithe map
• run \training-data\PatchifyTitheMask.ipynb to patchify the map's mask
• ~run \training-data\AddAugmentedPatches.ipynb to first filter out most of the ~ not made this yet

Training process:

• run \U-Net\train.ipynb to train the model on the patches

Land cover patch vectorisation:

• If you want to use the pre-trained model run \U-Net\predict.ipynb with model weights 'U-net-AG-D-ASPP-tithemaps.h5'
• run \post-processing\Predicted to predict and save masks for patches
  • *I need to make BATCH_SIZE scale relative to the number of input patches (maybe make BATCH_SIZE = (total input patches / 10))
• run \post-processing\full-map-from-patches.ipynb to stitch the patches into a full map prediction image and spatially project it using metadata from the reference map
• change stitching patches into full maps.ipynb to this name
• Using GRASS GIS processes (can be done more easily in QGIS with 'GRASS GIS provider' plug-in installed and activated) run the following tools to convert the binary raster prediction mask image into a set of vectorised polylines
  • r.mapcalc.simple
    • if(A>0,1,null())
    • input: {...}stitched_mask
    • output: {...}NoData_Set
  • r.thin
    • input: {...}NoData_Set
    • output: {...}thinned
  • r.to.vect
    • select line
    • input: {...}thinned
    • output: {...}polylines
  • simplify or v.generalize
    • input: {...}vectorised
    • output: {...}simplified
    • Simplification method: Douglas-Peucker Algorithm (Distance)
    • Tolerance: 0.5 m
• Visually inspect and fix errors
• Run vector-to-vector polygonisation to produce land cover patches.

Built on Ran and colleagues (2022) improved U-Net model for segmenting line elements from historic map documents.
Paper Citation: Ran, W. et al. (2022) ‘Raster Map Line Element Extraction Method Based on Improved U-Net Network’, ISPRS International Journal of Geo-Information, 11(8), p. 439. Available at: https://doi.org/10.3390/ijgi11080439. Link to GitHub: Raster-Map