# landtrendr-demo — the smallest chain that shows LandTrendr end to end. # # WHAT THIS MEASURES, AND WHAT IT DOES NOT # Measures: per-pixel NET change in annual median NBR (fitted last minus fitted first) # across 2015-2026, plus the start year of the largest-loss segment. Over the Oregon # Coast Range checkerboard the strong negatives are clearcuts. # Does NOT measure: burned area, forest loss in the Hansen/GFC sense, or anything # restricted to forest — there is no forest mask here. # # `magnitude` (band 0, what this recipe styles) is still NET change — but the note that # used to sit here, "the crate emits peak-to-trough as band 3; the engine wrapper slices # it off", is CLOSED as of 2026-08-19 (convergence LT-03). The engine now emits all four # kernel bands, [magnitude, year, rmse, peak_to_trough], and band 3 is the signed # peak-to-trough depth the disturbance literature thresholds on. It is measurably the # better detector: on Bootleg/MTBS through this same engine path, thresholding band 3 at # -0.20 scores recall 0.7772 / precision 0.9474 / kappa 0.6836 against band 0's # 0.5292 / 0.9631 / 0.4502. This demo keeps styling band 0 ON PURPOSE — a clearcut # checkerboard read over 12 years is a net-change picture, and the red/green two-sided # palette below needs a SIGNED band (peak-to-trough is non-positive by construction, so # it would render as one-sided). A recipe whose question is "where was there a # disturbance" should style/threshold band 3 instead. # # `year` dates disturbances ONE YEAR EARLY against LT-rs's own reference — that one is # still open, and it is the segment-START convention rather than an arithmetic bug # (docs/solutions/compute/landtrendr-rust-engine-e2e-findings.md; naming is convergence # decision D2). # # EXECUTION: runs on BOTH tiers. # * Rust engine: `folia run registry/recipes/landtrendr-demo` — measured green, see README.md. # * Browser: `registry/recipes/_runner/?recipe=landtrendr-demo` — the TS/WASM ops # (packages/compute/src/ops/{imagery-composite,landtrendr}.ts). The browser LandTrendr # is an independent port, NOT the LT-rs kernel the engine calls, so the two tiers are # not expected to be numerically identical. # # WHY IT EXISTS: derived from change-ensemble with the second detector and the ensemble # removed, so composite->landtrendr is SINGLE-CONSUMER and therefore fully quotable by the # gateway (multi-consumer chains are unpriced by design). # # KNOWN PLATFORM GAPS this recipe lives with (engine-side, worked around at recipe level only): # * No platform/collection filter reaches the grouped composite path (`query:` is threaded # only on the `scenes: true` arm), so Landsat 7 scenes cannot be excluded. A stray L7 # asset that answers HTML instead of a COG is a fatal parse, not a retry. Both were in # range on the measured runs and neither fired; the only lever if one does is the calendar. # * `max_segments: 8` with 12 annual bands is under the lt-rs F-test's `bands > 2*max_segments+1` # comfort zone (which wants >= 17). The ladder still walks down to a defined candidate and the # measured output is NOT degenerate (see README.md), but the knob is left as authored on # purpose: LandTrendr parameter adoption (LT-rs v0.4 + LT-GEE defaults) is a separate, # single-transaction change and must not be smuggled in here. name: landtrendr-demo version: "0.1" category: change-detection themes: [vegetation, forest, disturbance] family: forest-disturbance-landtrendr description: > Single-chain LandTrendr demo: annual NBR composites (cloud tier) feeding the LandTrendr temporal segmenter (client tier). Derived from change-ensemble with the second detector and ensemble removed so the composite->landtrendr chain is single-consumer and fully quotable. settings: composite_method: median limit: 6 # scenes PER YEAR (imagery_composite 3.0.0), not per window max_segments: 8 recovery_threshold: 0.25 inputs: aoi: bbox: [-123.57, 45.72, -123.10, 46.07] # Oregon Coast Range clearcut checkerboard crs_target: EPSG:32610 date_range: start: "2015-01-01" end: "2026-12-31" collection: # Canonical concept band codes only. The producer asset keys (nir08/swir22 on MSPC, # SR_B5/SR_B7 on LandsatLook) are the DISTRIBUTION's business, not the recipe's — # naming them here is what defect 60 was. uri: "@usgs/landsat/c2-l2" bands: [N, S2] layers: # Annual NBR. imagery_composite resolves the collection, fetches the [N, S2] pair per # year, and the band PAIR collapses to a normalized difference = NBR (the dataset owns # the SR scale/offset and the fill->NaN). compute/nbr-annual: type: raster uri: "${inputs.collection.uri}" params: bands: "${inputs.collection.bands}" bbox: "${inputs.aoi.bbox}" temporal: ["${inputs.date_range.start}", "${inputs.date_range.end}"] method: "${settings.composite_method}" group_by: year limit: "${settings.limit}" style: palette: rdylgn clim: [-0.3, 0.7] # Output bands (LT-03, 2026-08-19): FOUR, band-major — # b1 magnitude net change, fitted[last] - fitted[first] (signed; styled below) # b2 year START year of the largest-loss segment # b3 rmse observed-vs-fitted RMSE # b4 peak_to_trough signed disturbance DEPTH, non-positive by construction # `raster_calc` binds them positionally as b1..b4. detector/landtrendr: type: raster compute: op: landtrendr inputs: time_series: { layer: compute/nbr-annual } params: max_segments: "${settings.max_segments}" recovery_threshold: "${settings.recovery_threshold}" style: band: magnitude palette: rdylgn clim: [-0.5, 0.15] # loss red, stable yellow, recovery green views: - name: Change magnitude layers: [detector/landtrendr] - name: Annual NBR layers: [compute/nbr-annual] dashboard: patterns: # Year scrubber over the annual NBR cube: step the years the segmentation ran on. - slug: time-scrub-years slot: temporal layer: compute/nbr-annual # Click a pixel, see its NBR trajectory — the "show your work" half of the demo. - slug: pixel-trajectory-landtrendr slot: temporal layer: compute/nbr-annual element: folia-pixel-inspector license: recipe: Apache-2.0 output_dataset: CC-BY-4.0 upstream_data: - { source: landsat-c2-l2, license: public-domain }