defineTerrainSeg: defineTerrainSeg
In geodl: Geospatial Semantic Segmentation with Torch and Terra

defineTerrainSeg

R Documentation

defineTerrainSeg

Description

CNN-based semantic segmentation architecture of landform extraction or classification from a DTM.

Usage

defineTerrainSeg(
  segModel = "UNet",
  nCls = 2,
  cellSize = 1,
  spatDim = 640,
  tCrop = 64,
  doGP = FALSE,
  innerRadius = 2,
  outerRadius = 10,
  hsRadius = 50,
  smoothRadius = 11,
  actFunc = "lrelu",
  useAttn = FALSE,
  useSE = FALSE,
  useRes = TRUE,
  useASPP = TRUE,
  useDS = FALSE,
  pretrainedEncoder = TRUE,
  freezeEncoder = TRUE,
  avgImNetWeights = FALSE,
  enChn = c(16, 32, 64, 128),
  dcChn = c(128, 64, 32, 16),
  outChn = 64,
  btnChn = 256,
  dilChn = c(256, 256, 256, 256),
  dilRates = c(6, 12, 18),
  negative_slope = 0.01,
  seRatio = 8
)

Arguments

`segModel`	Segmentation architecture to use. Either UNet ("UNet"), UNet3+ ("UNet3p"), or UNet with a MobileNetv2 encoder.
`nCls`	Number of classes being differentiated. For a binary classification, this can be either 1 or 2. If 2, the problem is treated as a multiclass problem, and a multiclass loss metric should be used. Default is 2.
`cellSize`	Input resolution of DTM data. Default in 1 m.
`spatDim`	Input chip size. Default is 640 (640x640 cells)
`tCrop`	Number of rows and columns to crop from each side prior to passing LSPs to trainable model. Default is 64.
`doGP`	Whether or not to include Gaussian Pyramids of DTM and calulate LSPs at different scales. Default is FALSE. If FALSE, 6 LSPs are passed to model. If TRUE, 31 LSPs are passed to model.
`innerRadius`	Inner radius for annulus window for local TPI calculation. Default is 2 cells.
`outerRadius`	Outer radius for annulus window for local TPI calculation. Default is 10 cells.
`hsRadius`	Radius for circular moving window for hillslope TPI calculation. Defaults is 50 cells.
`smoothRadius`	Radius of circular moving window to smooth DTM prior to curvature calculations. Default is 11 cells.
`actFunc`	Defines activation function to use throughout the network when using UNet. "relu" = rectified linear unit (ReLU); "lrelu" = leaky ReLU; "swish" = swish. Default is "relu".
`useAttn`	TRUE or FALSE. Whether to add attention gates along the skip connections when using UNet, UNet with a MobileNet-V2 backbone, or UNet3+. Default is FALSE or no attention gates are added.
`useSE`	TRUE or FALSE. Whether or not to include squeeze and excitation modules in the encoder when using UNet. Default is FALSE or no squeeze and excitation modules are used.
`useRes`	TRUE or FALSE. Whether to include residual connections in the encoder, decoder, and bottleneck/ASPP module blocks when using UNet. Default is FALSE or no residual connections are included.
`useASPP`	TRUE or FALSE. Whether to use an ASPP module as the bottleneck as opposed to a double convolution operation when using UNet or UNet3+. Default is FALSE or the ASPP module is not used as the bottleneck.
`useDS`	TRUE or FALSE. Whether or not to use deep supervision when using UNet, Net with a MobileNet-V2 backbone, or UNet3+. If TRUE, four predictions are made, one at each decoder block resolution, and the predictions are returned as a list object containing the 4 predictions. If FALSE, only the final prediction at the original resolution is returned. Default is FALSE or deep supervision is not implemented.
`pretrainedEncoder`	TRUE or FALSE. Whether or not to initialized using pre-trained ImageNet weights for the MobileNet-v2 encoder. Default is TRUE.
`freezeEncoder`	TRUE or FALSE. Whether or not to freeze the encoder during training when using the MobileNet-V2 encoder. The default is TRUE. If TRUE, only the decoder component is trained.
`avgImNetWeights`	TRUE or FALSE. If three predictor variables are provided and ImageNet weights are used, whether or not to use the original weights or average them. This only applies when using MobileNet-V2 encoder. Default is FALSE.
`enChn`	Vector of 4 integers defining the number of output feature maps for each of the four encoder blocks for UNet or UNet3+. Default is 16, 32, 64, and 128.
`dcChn`	Vector of 4 or 5 integers defining the number of output feature maps for each of the 4 decoder blocks for UNet or UNet with a MobileNet-V2 encoder. Default is 128, 64, 32, and 16. Will need to change if using the MobileNet-V2 backbone.
`outChn`	Number of output channels for each decoder block for UNet3+. Default is 64.
`btnChn`	Number of output feature maps from the bottleneck block. Default is 256.
`dilChn`	Vector of 4 values specifying the number of channels to produce at each dilation rate within the ASPP module. Default is 256 for each dilation rate.
`dilRates`	Vector of 3 values specifying the dilation rates used in the ASPP module. Default is 6, 12, and 18.
`negative_slope`	If actFunc = "lrelu", specifies the negative slope term to use. Default is 0.01.
`seRatio`	Ratio to use in squeeze and excitation module when using UNet. The default is 8.

Details

Define a CNN-based semantic segmentation model for landform extraction or classification that includes a module that generates land surface parameters (LSPs) from the input DTM that are then passed to a semantic segmentation model. A UNet, UNet with a MobileNetv2 encoder, UNet3+, or HRNet architecture can be used. Gaussian pyramids can be calculated from the DTM to calculate LSPs at different scales. If Gaussian pyramids are not use, 6 LSPs are passed to the segmentation model. If Gaussian pyramids are used, 31 LSPs are passed to UNet3+. Model assumes a single band DTM of elevation measurements as input.