πŸ”§ Fix Forest monitoring Satellite imagery Forest Monitoring Β· step 20 Notebook By Β· Updated

Forest Loss Trend Not Comparable Over Time

On this page

Problem statement

The chart shows forest loss rising steeply after 2013, and the report says deforestation accelerated. Or it compares 2023 with 2005 and concludes loss has tripled. Or two analyses of the same region and year, a year apart, disagree. All three can come from treating an annual loss series as if every year had been measured the same way. Global Forest Change has been extended and reprocessed release by release, with improved inputs and algorithms; the providers themselves advise caution when comparing early and late parts of the record.

A trend in the data is a trend in the forest plus a trend in the measurement. Separating them means comparing like with like and checking against independent sources.

Quick answer

Fix the version, compare periods measured the same way, and cross-check with an independent series:

import numpy as np, rasterio
from rasterio.windows import from_bounds

W, S, E, N = -63.0, -10.8, -62.5, -10.3
PX_HA = 0.00025 * 111_320 * np.cos(np.radians(-10.55)) * 0.00025 * 110_574 / 1e4   # ~0.076 ha
def loss_series(v):
    with rasterio.open(f"/vsicurl/https://storage.googleapis.com/earthenginepartners-hansen/"
                       f"{v}/Hansen_{v}_lossyear_10S_070W.tif") as src:
        ly = src.read(1, window=from_bounds(W, S, E, N, src.transform))
    with rasterio.open(f"/vsicurl/https://storage.googleapis.com/earthenginepartners-hansen/"
                       f"{v}/Hansen_{v}_treecover2000_10S_070W.tif") as src:
        tc = src.read(1, window=from_bounds(W, S, E, N, src.transform))
    f = tc > 30
    last = int(v.split("-")[1]) - 2000
    return {2000 + y: (f & (ly == y)).sum() * PX_HA for y in range(1, last + 1)}

v12 = loss_series("GFC-2024-v1.12")
v11 = loss_series("GFC-2023-v1.11")
print(f"{'year':<6}{'v1.11':>9}{'v1.12':>9}")
for y in (2002, 2006, 2010, 2014, 2018, 2021, 2022, 2023, 2024):
    print(f"{y:<6}{v11.get(y, float('nan')):>9,.0f}{v12[y]:>9,.0f}")
year      v1.11    v1.12
2002      9,667    9,667
2006      7,110    7,110
2010        530      530
2014      1,047    1,047
2018      1,649    1,649
2021      1,741    1,741
2022      3,991    3,991
2023      1,240    1,240
2024        nan    1,645

Here the two releases agree exactly for every year they share. Where two releases give the same numbers for a year, that year was not reprocessed; where they differ, the older figure has been superseded. A trend assembled from figures quoted in different reports mixes releases.

Six small maps of the same RondΓ΄nia window, one per four-year period from 2001 to 2024, with the forest lost in that period in red: heavy scattered loss in 2001–2004, very little in 2009–2012, and a large clearing near the bottom in 2021–2024. Six small maps of the same RondΓ΄nia window, one per four-year period from 2001 to 2024, with the forest lost in that period in red: heavy scattered loss in 2001–2004, very little in 2009–2012, and a large clearing near the bottom in 2021–2024.
The same v1.12 loss layer as the table, in four-year periods. 2001–2004 cleared 25,764 ha across the whole window; 2009–2012 only 3,031; then it rose again, with one large block cleared near the southern edge in 2021–2024. A trend is a comparison between panels like these, and the guide’s point is that the panels must come from one release.

One release, release notes, consistent periods

1. Use one release for the whole series

Take every year from the same version. Do not append this year's figure from the new release to last year's series from the old one.

2. Read the release notes

Each GFC release documents what was reprocessed and which methodological changes apply to which years. Those boundaries are where a trend needs caution.

3. Compare within consistent periods

Compare averages of several years within a period measured the same way β€” for example two five-year periods after a method change β€” rather than single years across it.

4. Cross-check with an independent series

National monitoring systems (Brazil's PRODES for the Amazon), sample-based estimates, or a different product. Agreement in direction is reassuring; disagreement flags measurement change.

5. Report the version and the caveat

"Loss fell from X to Y (GFC-2024-v1.12; the providers caution against comparing years processed with different methods)" is honest. "Loss tripled" is not, if part of the tripling is the method.

Reasons an annual forest loss trend may not be comparable, with the check and fix for each. rule out the measurement firstwhat you seecheck β†’ fixnumbers differ between reportsdifferent releases β†’ one versiona step between yearsmethod change β†’ compare withinperiodslarge single-year swingsnoise β†’ multi-year meansa trend nobody else seescross-check PRODES or samplesCite the release and the caveat with every trend spatialworkflow.io
Before reading a trend, rule out the measurement.

Multi-year means and release differences

Example 1 β€” multi-year means instead of single years

def period_mean(s, a, b):
    return np.mean([s[y] for y in range(a, b + 1)])
for (a, b) in ((2001, 2005), (2006, 2010), (2011, 2015), (2016, 2020), (2021, 2024)):
    print(f"{a}–{b}: mean {period_mean(v12, a, b):7,.0f} ha/yr")
2001–2005: mean   5,665 ha/yr
2006–2010: mean   2,130 ha/yr
2011–2015: mean     894 ha/yr
2016–2020: mean   1,849 ha/yr
2021–2024: mean   2,154 ha/yr

Five-year means smooth the year-to-year noise and make changes of level visible: loss fell more than sixfold from 2001–05 to 2011–15, then more than doubled. They do not remove method changes, but they make them easier to spot as steps.

Example 2 β€” where the releases differ

diff = {y: v12[y] - v11[y] for y in v11}
changed = {y: d for y, d in diff.items() if abs(d) > 0.01 * max(v11[y], 1)}
print(f"years with >1% difference between releases: {sorted(changed) or 'none'}")
print(f"total 2001–2023: v1.11 {sum(v11.values()):,.0f} ha, "
      f"v1.12 {sum(v12[y] for y in v11):,.0f} ha")
years with >1% difference between releases: none
total 2001–2023: v1.11 59,660 ha, v1.12 59,660 ha

This is the check to run before combining figures from different reports: which years did the newer release change, and by how much?

Example 3 β€” the shape of the series against its own noise

vals = np.array([v12[y] for y in range(2001, 2025)])
diffs = np.diff(vals)
print(f"year-to-year change: median |Ξ”| {np.median(np.abs(diffs)):,.0f} ha, "
      f"largest {np.abs(diffs).max():,.0f} ha ({2002 + int(np.abs(diffs).argmax())})")
print(f"ratio of 2016–24 mean to 2001–09 mean: "
      f"{vals[15:].mean() / vals[:9].mean():.2f}")
year-to-year change: median |Ξ”| 854 ha, largest 6,049 ha (2007)
ratio of 2016–24 mean to 2001–09 mean: 0.46

In a small window, single-year swings are large relative to the mean. Statements about trends need several years on each side and the version caveat attached.

Mean annual forest loss in five periods from 2001 to 2024 in the RondΓ΄nia window. mean annual loss, ha2001–055,6652006–102,1302011–158942016–201,8492021–242,154030006000GFC-2024-v1.12 only: steadier than single years, still one release spatialworkflow.io
Five-year means: steadier than single years, still one release.

Why the method changed

Why the method changed

Global Forest Change was first published for 2000–2012 and then extended year by year. Later years benefit from better Landsat data (Landsat 8 from 2013), improved cloud masking and refined algorithms; some releases reprocessed parts of the record. The result is a better dataset overall and a less homogeneous time series.

Why single years are noisy

Annual loss depends on the weather (fire years), prices, policy and enforcement, and on how many clear observations the sensor got that year. In a small area, one large clearing or one cloudy year moves the annual figure substantially.

Why independent series help

National systems such as PRODES map clear-cut deforestation of primary forest with a consistent method and visual interpretation. They measure a narrower thing, but their consistency makes them a good check on the direction of trends in the global product.

Selected annual loss figures in GFC-2023-v1.11 and GFC-2024-v1.12 for the RondΓ΄nia window. two releases, year by yearyearv1.11v1.1220029,6679,667201053053020223,9913,99120231,2401,2402024not in release1,645Identical here for 2001–2023; the check takes one extra read spatialworkflow.io
Before combining figures from two reports, compare the releases they came from.

Fire years, recent years and primary forest

  • Fire years (2016, 2019–2020 in parts of the Amazon) raise loss through burning, not only clearing.
  • Recent years are the most likely to be revised in the next release.
  • Primary forest loss subsets are affected by the same method changes.
  • Area vs rate: a declining absolute loss can hide a rising rate on the remaining forest.
  • Regional differences: method changes affect regions differently, depending on cloudiness and forest type.

FAQ

Can I compare Hansen forest loss between 2005 and 2023?

With caution: methods improved over the record. Compare multi-year periods and cite the version.

Why do numbers for the same year differ between reports?

They used different releases; later releases reprocess recent years.

Should I mix figures from different GFC versions?

No. Take the whole series from one release.

How do I check whether a trend is real?

Compare with an independent series such as a national monitoring system, and look for steps at known method changes.

Why are annual figures so variable?

Weather, fire, policy and data availability all vary year to year; small areas amplify the noise.

Which version should I use?

The latest, consistently, with the version named in every figure.