Trends and Jumps in Thin DataCubes
tjdc-package.RdPixel level time-series analysis with the help of neighboring pixels.
Details
The `tjdc` package implements methods for fast trend and changepoint detection in stacked raster data, or data cubes.
Current version is based on the `stars`-package data management, but this might change in the future.
The main idea of the package is to provide user-friendly wrappers to split-calc-merge type of operations on high spatial dimension, low temporal dimension data-cubes. Motivation and research focus has been on datacubes for the Finnish National Forest Inventory with spatial resolution of 16m x 16m but with mere 10-15 time steps. Such series are below the lower threshold for sophisticated change point and trend analysis techniques, but assuming spatial correlation between nearby pixels we might be able to improve the power.
The statistical functionality is roughly in two groups.
Trend analysis
This is functionality carried over from the `ConMK` package. The main ideas revolves around the contextual Mann-Kendall's test [tj_contextual_mann_kendall()]. Nothing new in this package compared to the `ConMK` other than switching the retired `raster` package to `stars`.
Change-point analysis
For short data sequences with high noise, detecting reliably even at most one change-point is challenging. A simple linear trend model with a possible jump is estimated so that the jump events are correlated between neighbours.