The workflow used in this worksheet emerges from a simple constraint: spatial systems such as the Burgwald are too complex, and too heterogeneous in scale, to be optimised directly for gauge placement or model validation. If structural variability, segmentation scale, and process signals are mixed into a single step, the result is usually unstable, hard to interpret, and difficult to reproduce. One possible framework therefore separates four stages: first, structural stratification on coarse tiles using information-theoretic and compositional metrics; second, adaptive segmentation on representative test tiles to discover meaningful spatial scales; third, transfer of the chosen segmentation configuration to the full domain and aggregation into physiographic process strata; and only then, fourth, the use of these strata as a basis for network optimisation and hydrological reasoning. Structural, physiographic, and contextual information are treated as different “layers of evidence” rather than one big predictor stack. This staged reduction of spatial complexity is designed to keep scale decisions explicit, to link segments back to physical processes, and to provide a transparent foundation for later decisions such as station placement or sampling design.
Goals
- Explicitly reflect on where this workflow constrains thinking,
and where alternative structuring choices might be equally valid or preferable. - Use the existing spatial data pipeline as a conceptual technical base.
- Understand the difference between structural, physiographic, and process-oriented stratification.
- Develop a multi-stage workflow that:
- first reduces spatial complexity structurally,
- then discovers suitable segmentation scales,
- then builds process strata,
- and only afterwards moves toward optimisation.
- Learn how to combine:
- information-theoretic and structural metrics (tile level),
- terrain- and land-surface descriptors (segment level),
- adaptive spatial segmentation (test tiles → full AOI).
- Build a reproducible basis for later:
- station placement,
- sampling design,
- or model evaluation.
You may use ChatGPT only to improve wording and structure of your own notes.
- Write your own bullet points first.
- Use ChatGPT only to shorten, clarify, or rephrase.
- Do not ask ChatGPT for full solutions or methodological decisions.
- If unsure, explicitly write “I am unsure about …” — ChatGPT may then give guiding questions, not answers.
- Keep answers short and technical.
Prerequisites
Before starting, make sure that:
- The project repository is cloned and opened as an RStudio Project.
- The basic setup script has been executed once:
- project paths are defined,
- required input folders exist.
- The following spatial data are available for the Burgwald AOI:
- a categorical land-cover / Sentinel classification raster,
- a digital elevation model (DEM),
- a polygon AOI defining the study region.
- Required R packages are installed and load without errors.
The specific choices shown in the next tasks reflect the instructor’s implementation.
Other choices are desireable if they are conceptually consistent and well argued especially within the individual project approach.
Task 1 – Base data and spatial units (Step 0 in the diagram)
Clarify which spatial units all later steps will build on.
- Inspect spatial resolution and extent of:
- the land-cover raster,
- the DEM.
- Define one coarse structural unit for Step 1:
- e.g. a fixed grid of tiles (e.g. 2×2 km).
- Define one adaptive unit candidate for Steps 2a/2b:
- e.g. image segments, supercells, or similar.
Short written output:
- Which spatial unit is used first, and why is it suitable as a structural base?
- Which processes are not well represented by fixed grid cells (e.g. cold-air drainage, flow paths)?
- Why might adaptive units be preferable for later process-oriented analysis?
Task 2 – Structural pre-stratification on tiles (Step 1: structural strata)
Reduce spatial complexity on the level of coarse tiles (structural, not yet process-based).
- For each coarse tile, compute structural descriptors, e.g.:
- land-cover thematic diversity,
- simple configurational / composition metrics (e.g. forest fraction, entropy).
- Cluster the tiles based on these structural descriptors into structural tile clusters.
Short written output:
- Which variables are used, and why are they considered structural (not process variables)?
- What does tile clustering achieve at this stage (in terms of complexity reduction)?
- Why is this not yet a physiographic or process-based stratification?
Task 3 – Selection of representative test tiles (end of Step 1)
Restrict the more expensive segmentation experiments to a small set of test tiles.
- From each structural tile cluster, select one or more representative tiles, e.g.:
- closest to the cluster centroid in descriptor space,
- or with median values for key structural metrics.
- Document the selection rule so that it is reproducible.
Short written output:
- Why is it useful to restrict further analysis to representative test tiles?
- What are the risks if segmentation is tuned on the full AOI without this step?
- How does this step improve reproducibility and efficiency of later segmentation tuning?
Task 4 – Adaptive segmentation on test tiles (Step 2a: segmentation as scale discovery)
Now explore how the landscape self-organises into segments on the representative tiles.
- Choose one segmentation approach:
- e.g. supercells, mean-shift, region growing, or similar.
- Apply it only to the selected test tiles.
- Systematically vary spatial and/or spectral parameters, e.g.:
- bandwidths, compactness, scale parameters.
Use raw or low-level variables for this step:
- e.g. DEM, land-cover, basic reflectance channels.
- Avoid using derived process variables here.
Short written output:
- Which input variables are used for segmentation on test tiles?
- Why should strongly derived variables (e.g. complex indices, stability classes) be avoided at this stage?
- What does “over-segmentation” vs. “under-segmentation” mean in this context?
Task 5 – Objective evaluation of segmentation quality (Step 2a: choose scales)
Evaluate candidate segmentations on the test tiles using quantitative criteria.
Define at least three evaluation criteria, such as:
- Segment size distribution (too many tiny segments vs. too few large segments),
- Intra-segment homogeneity (e.g. variance of elevation or land-cover inside segments),
- Inter-segment contrast (difference between neighbouring segments),
- Optional: stability under small parameter changes (segments change only moderately).
Apply these criteria to the test tiles for several parameter settings.
Short written output:
- Which metrics indicate good segmentation behaviour for the Burgwald use case?
- Why is there no single “optimal” segmentation parameter set?
- What does a Pareto-optimal solution mean here (trade-offs between criteria)?
Task 6 – Segmentation transfer and physiographic strata (Steps 2b and 3)
Move from test-tile experiments to wall-to-wall segments and then to physiographic strata. This step formalises one interpretation of how segments relate to dominant processes;
other process groupings are possible and may be more appropriate for different research questions.
Segmentation transfer (Step 2b)
- Fix one or a small set of segmentation configurations (method, scale, features) based on Task 5.
- Apply the chosen configuration(s) to the full Burgwald AOI.
- Result: a continuous segmentation map of spatial analysis units.
Physiographic / process stratification (Step 3)
For each segment in the full AOI, attach process-relevant attributes, e.g.:- terrain position (valley, slope, ridge),
- mean elevation,
- forest / vegetation fractions or surface fractions.
Cluster segments into physiographic process strata, e.g.:
- cold-air drainage strata,
- orographically exposed strata,
- forest–atmosphere coupling strata.
Short written output:
- Which variables now directly relate to physical processes (not just structure)?
- Why should collinear or strongly derived variables be excluded from clustering?
- How do these physiographic strata differ from the initial structural tile clusters?
Task 7 – Contextual annotation for later decision making (link to Step 4)
Prepare for later network optimisation / hydrological design without mixing decision criteria into the clustering space.
- Attach contextual descriptors to each segment or stratum, e.g.:
- exposure/windwardness (annotation only),
- accessibility / infrastructure proximity,
- existing stations, land ownership (if available).
- Keep these descriptors separate from the variables used for clustering into physiographic strata.
Short written output:
- Why are these variables treated as annotations rather than clustering dimensions?
- How can they still influence final site selection (e.g. station placement)?
- What problems arise if such contextual variables are included directly in clustering?
Take-home – From structure to robust design (Steps 1–4)
Write one short paragraph (max. 6–7 sentences) that positions this workflow::
- why spatial complexity in the Burgwald must be reduced in stages:
- from structural tiles (Step 1),
- via segmentation scale discovery and transfer (Steps 2a/2b),
- to physiographic process strata (Step 3),
- how structural, physiographic, and contextual/annotative information differ,
- why segmentation should precede process-based clustering,
- and how this workflow supports robust station or sampling design and later steps such as:
- optimisation of gauge locations,
- uncertainty reduction,
- or hydrological validation (Step 4).