# The Forest Remembers: Sources and Methods Edition: 15 September 2026. The satellite analysis covers Germany, Austria, Czechia, Poland, metropolitan France and mainland/Balearic Spain, 1985-2024. Historical reconstruction, official forest area, mapped canopy disturbance and manual-reference disagreement remain separate evidence classes. ## Satellite Source and Acquisition The European Forest Disturbance Atlas (EFDA) v3.0.0 was released on 13 September 2026 at , under CC BY 4.0. Every annual observation comes from that release. The v3 detector uses deep learning; the 2025 method paper documents the forest mask and agent framework but does not validate the new v3 detection algorithm. Seven native products are used: forest-land-use mask, annual disturbance, annual agent, aggregated agent, annual severity, native disturbance count and latest disturbance year. The probability product is not acquired. No native greatest-disturbance-year product exists in v3. Acquisition reads ZIP central directories through exact HTTP byte ranges. It accepts only the pinned record and named products, at the archive root or in its matching country directory. Each selected member's length, CRC32 and SHA-256 is verified; cached bytes are fully rechecked. The complete-ZIP MD5 is recorded but not claimed as verified because the complete ZIP is not downloaded. Full archives total about 20 GB; selected products require roughly 3 GB transferred and 5 GB as native rasters. Raw rasters stay outside public deployment. ## Native Counting and Recurrence All pixels are read in bounded 512-by-512 windows, without sparse sampling. The fixed forest mask equals 1. Each 30-by-30-meter EPSG:3035 pixel represents 0.09 hectares. Annual disturbance is the source value 1; 255 is not a forest denominator. Annual share is disturbed pixels divided by mapped forest pixels. France has 97 and Spain 154 source pixel-year disturbance indicators outside the fixed forest mask. They are recorded diagnostically and excluded before all forest calculations. They are not unique-area estimates. The other four countries have zero such indicators; within-mask count and spatial checks pass. Four agent classes reconcile to all detected hectares. Wind and bark beetles remain combined; unclassified events stay in the denominator. Harvest and salvage-related canopy changes are included. These are not measures of permanent deforestation or natural-disaster area. The native terminal event count and latest-year layer must agree with the annual sequence at every forest pixel. Ever-disturbed and repeatedly disturbed footprints use the same fixed denominator and 1985-2024 window. Repeated area is a subset of ever-disturbed area: their association is partly mathematical, not evidence of failed recovery. Annual recurrence uses previous events only. A fixed alternative looks back ten complete years, excluding the current year; it is unavailable before 1995. Pooled conditional rates use current disturbed hectares as weights. Their pre-period is 1995-2017, not an early period with a shorter lookback. The 2024 percentile is `100 * (earlier values below + 0.5 * ties) / 39`, using 1985-2023. Period comparisons use annual means, not cumulative area over unequal periods. The 2018 split is a descriptive episode boundary, not an estimated change point. Baseline starts of 1985, 1990, 1995 and 2000 are supplied. No weather regression, causal attribution or structural-break test was estimated. ## Severity and Timing Signed dNBR is summarized only among events with valid severity. Mean and 75th percentile are distribution summaries, not confidence limits. A value of 0.3 does not mean 30 percent tree mortality, biomass loss or financial damage. Source nodata is excluded; negative values remain. Metadata documents -1 to 1. Values outside it but inside the physical -2 to 2 range are retained, counted and accompanied by a restricted-range sensitivity. Values outside the physical range fail aggregation. Exact source-value histograms supply linearly interpolated quantiles. Greatest-disturbance year is derived by maximizing valid annual dNBR over detected events, with the earliest year winning exact ties. Pixels without valid event severity remain unassigned and are counted. Latest and greatest-year distributions are separate in country summaries and annual exact values. ## The 2024 Spatial Grid Every native forest pixel enters one ten-kilometer EPSG:3035 cell. Forest and disturbed counts reconcile exactly to national totals. Polygons transform to geographic coordinates for display; colors use source counts, not screen area. Country shares are ratios of sums, not averages of cell percentages. Cells with mapped forest and zero disturbance remain. Empty locations are not assumed to contain undisturbed forest. Small forest denominators can create high cell rates, so both hectares and shares are exposed. The map stays fixed at 2024. ## Manual-Reference Disagreement The cleaned reference is v2, released 25 August 2026 at , CC BY 4.0. The 126,187,649-byte CSV has publisher MD5 `923cf9a52d2fcd49c9d1b30d2217bb89`; the cleaning R code has MD5 `a67dbeb2b95d5d2932033046b37247e9`. Lengths and hashes are checked on download. Each country has 1,000 sites: half from an older disturbed map stratum and half from an undisturbed stratum. Sites outside the v3 mask or extent are separately reported and excluded from within-mask comparisons. Labels end in 2023, giving 39 annual comparisons per included location. Equal strata are not national prevalence weights. These are unweighted sample diagnostics, not national accuracy estimates or error-corrected hectares. Independence from v3 training has not been established. Exact-year TP, FP, FN and TN are reported by country, year and stratum. Commission is FP / mapped events; omission is FN / reference events. Stable site-years dominate overall agreement. The one-year-tolerance matcher is one-to-one, maximizing pairs, then minimizing absolute timing distance, then choosing the earliest lexicographic pairing. Exhaustive small-case tests check the dynamic program. Offset counts and unmatched events remain available. The high-confidence flag partly uses Landsat spectral information and is not independent confirmation. There are no manual 2024 observations or causal-agent labels. Public diagnostics retain raster definitions but omit execution paths. ## Official Forest-Area Sensitivity Eurostat `for_area_efa`, closing stock `STK_CL`, forest `FOR`, thousand hectares `THS_HA`, 2023, supplies the alternative forest denominator. It excludes other wooded land and national nonforest area. The frozen response and derived hectares remain separate. France is metropolitan, including Corsica, and flagged estimated/provisional (`ep`). Spain is imputed (`i`); its national total includes the Canary Islands, outside the EFDA extent. This fixed 2023 denominator is a scale check, not annual forest at risk from 1985 onward. It changes rate levels but cannot change a country's year rankings or post/pre ratios. Those identities do not resolve differing definitions. ## Longer Historical Clocks Pollen windows come from Githumbi et al. (2022), , and LandClimII at . The full reconstruction uses 1,128 pollen records across 11,700 years. This display retains six windows and a union frame of 392 one-degree locations, not complete coverage in every period. Woody cover is evergreen plus summer-green vegetation, including its source woody taxa; it is not FAO-defined forest area. All-zero source rows are missing. Quality 1 requires a large lake or multiple sites; quality 2 uses one small lake or bog. Component standard errors remain separate because covariance is absent. The recent window is nominally 1850-1950 in the spreadsheet, while the paper describes the upper boundary as the coring year. It is never a 2024 inventory. Recovered historical derivatives were checked against their preserved receipt and hashes. This edition did not re-extract the workbook: PANGAEA retrieval was unavailable. Archive/workbook hashes and the existing `build_history.py` owner remain documented. This is a reproducibility limitation, not a fresh extraction claim. Natural Earth's public-domain modern borders provide orientation only. Modern forest area comes from the corrected May 2026 State of Europe's Forests 2025, Table 1.1-3: 1990/2000/2010/2020/2025 for EU-27 and its wider reporting region. No annual observations are interpolated and it is not joined to pollen cover. FAO definitions retain temporarily unstocked forest expected to regenerate. ## Reproduce from the Repository Use Python 3.12 and pinned raster requirements on a host that permits GDAL. Commands are relative to the repository root; concurrency is bounded: ```sh python -m pip install -r scripts/forest-regime/efda/requirements.txt python scripts/forest-regime/efda/test_acquire.py python scripts/forest-regime/efda/test_aggregate.py python scripts/forest-regime/efda/validate_reference.py --self-test python scripts/forest-regime/efda/acquire_efda_v3.py de at cz pl fr es --workers 1 --data-dir work python scripts/forest-regime/efda/aggregate_efda_v3.py de at cz pl fr es --workers 1 --input-dir work/raw --output-dir work/review python scripts/forest-regime/efda/download_reference.py --output-dir work/reference python scripts/forest-regime/efda/validate_reference.py --reference work/reference/disturbance_database_clean.csv --raw-root work/raw --output-dir work/review --countries Germany Austria Czechia Poland France Spain python scripts/forest-regime/publish_data.py --processed-dir work/review python scripts/forest-regime/build_article.py ``` The publisher independently reconciles native counts, hectares, annual shares, agent totals, terminal distributions, percentiles, map cells and manual counts before writing any output. Its manifest binds source hashes to publication bytes. No pixel is adjusted manually for a chart. ## Rights and Interpretation Upper Spreewald: A.Savin, Wikipedia, Free Art License 1.3. Wurmberg/Harz: Llez, CC BY-SA 3.0. Exact source pages, dates, original/export hashes and resizing changes are in the photo receipt. They show different places in July 2024, not before/after states, and do not validate individual pixels. Historical and economic context is sourced in the article. Mechanisms involving salvage, future supply, carbon and recovery are interpretations bounded by those sources. No causal coefficient, monetary damage total, forecast of collapse or 2024 manual validation is claimed.