DEM_segmentation

In this repo, we implement a Unet model with data preprocessing for segmenting geographical feature such as waterways on Ditial Elevation Maps (DEM).

Setup

Prerequisites

• Tensorflow

• keras

• gdal (optional, matters when you need to retile from a large DEM)

  • Anaconda: conda install -c conda-forge gdal
  • pip : pip install GDAL

• Clone this repo:

git clone git@github.com:Chen-Yifan/DEM_segmentation.git
cd DEM_segmentation

Generate Dataset from DEM

Original TIFF files:

DEM: mclearn_roi.tif

Annotated Mask:

annotated_mask.tif where 1 indicates annotated, 0 not

Binary Mask:

cl1.tif -- 5 m buffer around the features  with 1 means having feature there
10m_buffer_3443.tif -- 10 m buffer around the features 

1. Merge to 6 bands

Combine the separate bands in a single image; all bands will be initialized using 0

gdal_merge.py -separate -init 0 -o mclean_roi_merge.tif mclean_roi_slope.tif mclean_roi_aspect.tif mclean_roi_rough.tif mclean_roi_tpi.tif mclean_roi_tri.tif mclearn_roi.tif

2. Create Tiled Images (Ex: Size 128x128 with overlap 64)

gdal_retile.py -v -r near -ps 128 128 -co “TILED=YES”  -targetDir frames_128_overlap  -tileIndex  tiles_frames  -overlap 64     -csv frames.csv  -csvDelim ,  mclean_roi_merge.tif 

gdal_retile.py -v -r near -ps 128 128 -co “TILED=YES”  -targetDir masks_128_overlap  -tileIndex  tiles_frames  -overlap 64     -csv masks.csv  -csvDelim ,  cl1.tif 

(or 10m_buffer_3443.tif for 10 m buffered feature)

gdal_retile.py -v -r near -ps 128 128 -co “TILED=YES”  -targetDir annotations_128_overlap  -tileIndex  tiles_annotations  -overlap 64     -csv annotations.csv  -csvDelim ,  annotated_mask.tif 

Model Setup

Based on Unet, we add Convolution2D regularization parameter, Dropout layer, and we also modify the number of filters. We use Adam algorithm as our optimizer.

Train and Test

Train_frame_path and train_mask_path contains npy tile files. To train the model: ./scripts/train.sh

main.py: the main function

data_loader.py: load data as arrays, and do preprocess such as normalization. In this part, we only experiment with single-band DEM input or its gradient. To include their derivatives, please change the shape of the arrays also the size of input, output of the model.

dataGenerator.py: data batch augmentation for training.

build_model.py: the actual model structure and compile the model.

define_model.py: fit the model and callback specifications for train and test.

losses.py: loss functions that might be helpful.

metrics.py: metrics to measure the performance. Please use iou_label during training, and feel free to test your model after training is done with dissimilarity_loss