RStoolbox (version 0.3.0)

pifMatch: Pseudo-Invariant Features based Image Matching

Description

Match one scene to another based on linear regression of pseudo-invariant features (PIF).

Usage

pifMatch(
  img,
  ref,
  method = "cor",
  quantile = 0.95,
  returnPifMap = TRUE,
  returnSimMap = TRUE,
  returnModels = FALSE
)

Arguments

img

RasterStack or RasterBrick. Image to be adjusted.

ref

RasterStack or RasterBrick. Reference image.

method

Method to calculate pixel similarity. Options: euclidean distance ('ed'), spectral angle ('sam') or pearson correlation coefficient ('cor').

quantile

Numeric. Threshold quantile used to identify PIFs

returnPifMap

Logical. Return a binary raster map ot pixels which were identified as pesudo-invariant features.

returnSimMap

Logical. Return the similarity map as well

returnModels

Logical. Return the linear models along with the adjusted image.

Value

Returns a List with the adjusted image and intermediate products (if requested). #'

• img: the adjusted image

• simMap: pixel-wise similarity map (if returnSimMap = TRUE)

• pifMap: binary map of pixels selected as pseudo-invariant features (if returnPifMap = TRUE)

• models: list of linear models; one per layer (if returnModels = TRUE)

Details

The function consists of three main steps: First, it calculates pixel-wise similarity between the two rasters and identifies pseudo-invariant pixels based on a similarity threshold. In the second step the values of the pseudo-invariant pixels are regressed against each other in a linear model for each layer. Finally the linear models are applied to all pixels in the img, thereby matching it to the reference scene.

Pixel-wise similarity can be calculated using one of three methods: euclidean distance (method = "ed"), spectral angle ("sam") or pearsons correlation coefficient ("cor"). The threshold is defined as a similarity quantile. Setting quantile=0.95 will select all pixels with a similarity above the 95% quantile as pseudo-invariant features.

Model fitting is performed with simple linear models (lm); fitting one model per layer.

Examples

# NOT RUN {
library(raster)

## Import Landsat example data
data(lsat)

## Create fake example data
## In practice this would be an image from another acquisition date
lsat_b <- log(lsat)  

## Run pifMatch and return similarity layer, invariant features mask and models
lsat_b_adj <- pifMatch(lsat_b, lsat, returnPifMap = TRUE, 
                         returnSimMap = TRUE, returnModels = TRUE)
# }
# NOT RUN {
## Pixelwise similarity
ggR(lsat_b_adj$simMap, geom_raster = TRUE)

## Pesudo invariant feature mask 
ggR(lsat_b_adj$pifMap)

## Histograms of changes
par(mfrow=c(1,3))
hist(lsat_b[[1]], main = "lsat_b")
hist(lsat[[1]], main = "reference")
hist(lsat_b_adj$img[[1]], main = "lsat_b adjusted")

## Model summary for first band
summary(lsat_b_adj$models[[1]])
# }