Forest gain and loss: GuidosToolbox's change of a forest

The first three parts of this series compared maps of many classes: a transition matrix, its intensity, its agreement with a reference. Conservation questions are often simpler and sharper: where did the forest go, where did it come back, and is what remains still forest interior? In this last part you will look at one class, the forest, with the change analyses of GuidosToolbox: the simple change (gain, loss and persistence of the forest), the morphological change, which tells change that altered the forest's structure from the shifts of its edges that two maps of the same forest always show, and the change of its fragmentation at several scales. You will run them on Sibuyan's forest from 1997 to 2011 (the data of part 1), on Plum Island's from 1985 to 1999, and on a map moved by one cell, to see what a misregistration does. At the end you will be able to report how much forest was lost, how much of that loss you can trust, and how the forest left behind was changed.

The idea#

From many classes to one#

A transition matrix treats every class alike. GuidosToolbox (Vogt & Riitters, 2017), the image analysis software of the European Commission's Joint Research Centre, looks at one class at a time, the foreground (here the forest), against everything else, the background. Its maps are coded 0 for missing cells, 1 for background and 2 for foreground; in Terraflux every operator of this page takes instead a foreground setting that names the classes of your map that are forest (foreground = [1] for Sibuyan's class 1), and treats cells without a value (the sea) as missing.

With two binary maps, every cell is one of four things:

(255 marks a cell missing at either date.) This is GuidosToolbox's simple change, change.simple. Its loss is the forest row of the transition matrix of part 1 summed over every other class; its gain, the forest column.

Not all change on a map is change on the ground#

Two maps of the same forest, made from two images, never align perfectly: the images are registered to within a fraction of a cell, the classifier hesitates on mixed cells, and the forest's edge is a fuzzy line. Every one of these errors shows up as change along the edges of the forest, one or two cells wide. Townshend et al. (1992) showed how much change a registration error of a fraction of a cell creates in satellite images, and Olofsson et al. (2014) why map errors make change maps the least accurate of all land maps.

Morphological change detection (MCD; Seebach et al., 2013), change.morphological, separates the change that alters the structure of the forest from that edge noise:

Enlarge

A clearing is structural only if, somewhere inside it, a 5 × 5 block went from all forest to all background. That makes MCD's answer depend on the size of the cells: at 250 m a seed is a block of 1.25 km a side, at 30 m one of 150 m.

What the change did to the forest that remains#

Losing 10% of a forest from its edge and losing it as a hundred holes in its middle are different events for the species of its interior. GuidosToolbox measures the forest area density (FAD: the share of forest in a window around each forest cell; Riitters et al., 2000) at both dates and asks how the forest present at both dates, the common forest, moved between its classes (rare, patchy, transitional, dominant, interior, intact). The same fragmentation measured cell by cell at one window is FOS (fixed observation scale). The fragmentation series covers both in depth; here they answer one question of change detection: did the forest that was not cleared lose its surroundings?

When to use it#

• Deforestation and regrowth maps for a report or a paper: the area lost and gained (simple change), and how much of it is structural (morphological change), from two forest maps, global products such as tree cover maps, or two classifications of your own.
• Checking a pair of maps before trusting their change. A large share of spurious change says the maps differ mostly along the edges, the signature of misregistration or of classifiers that disagree on mixed cells.
• Forest interior and fragmentation indicators: whether the forest that remains is less interior than before (Riitters & Wickham, 2012; Riitters et al., 2016).
• Comparing a scenario with the present: the second date can be a simulated map (see the land change category).

It is not for many classes at once (use the transition matrix of part 1); nor for the rate of change against the time (the intensity analysis). And morphological change does not correct a map: spurious change is change MCD cannot confirm, not change it has shown to be false. For the area of change with confidence intervals, a sample of reference observations is the proper tool (part 3).

The habitat patterns series runs simple and morphological change next to MSPA on a large forest; the fragmentation series runs FAD's and FOS's change at five scales and writes GuidosToolbox's reports. This part is the change-detection view: what these numbers add to a transition matrix, and when to believe them.

What you need#

• Two maps of the same grid: same extent, cell size and coordinate system, made the same way. Any classes: the foreground setting says which are forest.
• Missing cells as no data (the sea, clouds): they are left out (code 255).
• Here, the file of the series, sibuyan_landuse.tif (108 × 128 cells of 250 m, bands 1997 and 2011, forest class 1; see part 1), and Plum Island's maps of the CLUE-S series, pie_landuse.tif (434 × 497 cells of about 100 m, bands 1985, 1991 and 1999, forest class 1).

Build the model#

The template#

format = "terraflux/1"

[[nodes]]
id = "before"
op = "nlm.random_cluster"
params = { rows = 250, columns = 250, p = 0.55, weights = [0.4, 0.6], seed = 27 }
note = "The forest at the first date, drawn by chance: 1 the background, 2 the forest (in its place, a raster.read of a map of yours)."

[[nodes]]
id = "pressure"
op = "nlm.gaussian_field"
params = { rows = 250, columns = 250, range = 30, seed = 28 }
note = "Where clearing presses, from 0 to 1."

[[nodes]]
id = "after"
op = "raster.calc"
inputs = { f = "before.map", p = "pressure.map" }
params = { expression = "1 if f == 2 and p > 0.6 and nbsum(f == 1, 1) >= 1 else 2 if f == 1 and p < 0.2 and nbsum(f == 2, 1) >= 5 else f" }
note = "The second date: the forest cleared from its edges where the pressure is high, regrown in the gaps where it is low."

[[nodes]]
id = "simple"
op = "change.simple"
inputs = { before = "before.map", after = "after.map" }
note = "Every cell's change: stable background, gain, loss, stable forest; stats counts them."

[[nodes]]
id = "morphological"
op = "change.morphological"
inputs = { before = "before.map", after = "after.map" }
note = "The same change with the spurious one at the forest's edges (a cell's shift) told apart; stats has the net change and the elasticity."

[[nodes]]
id = "fad"
op = "change.fad"
inputs = { before = "before.map", after = "after.map" }
params = { foreground = 2, windows = [7, 13, 27, 81] }
note = "FAD at four scales at both dates: how many forest cells moved to a more fragmented class or a less fragmented one (stats), and the matrix of the classes."

[[plots]]
id = "plot_simple"
title = "Simple change"
layers = [{ from = "simple.map" }]

[[plots]]
id = "plot_morphological"
title = "Morphological change"
layers = [{ from = "morphological.map" }]

1. before (nlm.random_cluster) draws the first date: clusters of two classes, 1 background (40%) and 2 forest (60%), already in GuidosToolbox's coding, so no foreground is needed downstream.
2. pressure (nlm.gaussian_field) draws a smooth surface from 0 to 1: where clearing presses.
3. after (raster.calc) makes the second date from the first: a forest cell with at least one background cell among its 8 neighbours (nbsum(f == 1, 1) >= 1) is cleared where the pressure passes 0.6, and a background cell with at least 5 forest neighbours regrows where it is below 0.2. Only the edge retreats, by one cell: keep this in mind when reading the morphological change below.
4. simple (change.simple) takes the first date on before and the second on after. Swap them and gain and loss swap too.
5. morphological (change.morphological) takes the same two inputs.
6. fad (change.fad) also takes the two forest maps and computes FAD at both dates itself, at the windows given (sides in cells, odd).

In the application, add each change node from the library's subject Change detection, drag before.map onto its before port and after.map onto after, and select the node to see its settings in the Node pane.

First date

Second date

On your maps: Sibuyan's forest#

format = "terraflux/1"

[[nodes]]
id = "before"
op = "raster.read"
params = { path = "sibuyan_landuse.tif", band = "1997" }
note = "Land use in 1997: 1 forest, 2 coconut, 3 grassland, 4 rice, 5 other."

[[nodes]]
id = "after"
op = "raster.read"
params = { path = "sibuyan_landuse.tif", band = "2011" }

[[nodes]]
id = "simple"
op = "change.simple"
inputs = { before = "before.raster", after = "after.raster" }
params = { foreground = [1] }

[[nodes]]
id = "morphological"
op = "change.morphological"
inputs = { before = "before.raster", after = "after.raster" }
params = { foreground = [1] }

[[nodes]]
id = "fad"
op = "change.fad"
inputs = { before = "before.raster", after = "after.raster" }
params = { foreground = [1], windows = [7, 13, 27] }

Nodes 1 and 2 read the two bands of the file by their names. Each change node gets foreground = [1]: forest is class 1, and coconut, grassland, rice and other land are the background. To call coconut plantations forest too, write foreground = [1, 2]: the answer changes, and that choice belongs in the methods of a paper. The windows of FAD are in cells: 7, 13 and 27 cells of 250 m are 1.75, 3.25 and 6.75 km.

Read the results#

Simple change: how much forest#

Enlarge

The stats table: 4,160 cells of stable forest, 563 of loss (35.2 km²), 0 of gain, 2,503 of stable background, and 6,598 missing (the sea). These are exactly the numbers of the forest's row and column in the transition matrix of part 1: 320 cells to grassland, 126 to coconut, 116 to rice, 1 to other; nothing into forest. Simple change adds no information to the matrix; it maps it in one class and in GuidosToolbox's codes, which the other analyses read.

The losses lie along the forest's northern and south-western edges, where the farmland and grassland of the coast met it; part 1's map of transitions shows which class took each cell.

Morphological change: how much of it is structural#

Enlarge

Two thirds of the loss (375 of 563 cells) is spurious: strips one or two cells wide along the forest's edge. Of the 208 cells of structural loss, 188 were cleared, 19 were forest at both dates and 1 background at both: the 3 × 3 closing of step 3 fills the small islands and gaps inside a clearing, so the map shows where the forest's structure was lost rather than the exact cells cleared. Report the area lost from simple change, and the structural part from morphological change.

Enlarge

Is the spurious change real? On these maps, probably much of it is: a forest frontier advancing by a cell (250 m) in fourteen years is plausible. MCD cannot know. It says that this change could have been made by the two maps being shifted or classified differently along the edge, and that the structural part could not. The elasticity of 1.05 says the forest's interior shrank about as fast as the forest: Sibuyan lost forest from its outside, not from within.

A misregistration: change where nothing changed#

To see what MCD is for, compare the forest of 2011 with itself, moved one cell (250 m) to the east, as two images registered one cell apart would be. Nothing changed on the ground.

Simple change

Morphological change

Enlarge

Simple change reports 354 changed cells, 8.5% of the forest, more than half the real loss of fourteen years, from a shift of one cell. Morphological change reports none. When two maps come from different sensors, dates of the year or classifiers, run morphological change first: a large share of spurious change is a warning about the maps before it is a finding about the forest.

Resolution matters: Plum Island#

Plum Island's forest (1985 to 1999, cells of about 100 m) both lost and gained: simple change gives 4,906 cells of loss and 1,270 of gain, a net loss of 3,636 cells (7.4% of the forest). Morphological change calls all 6,176 changed cells spurious: no structural loss, no structural gain.

Enlarge

The change here is suburban growth in many small patches (part 1 of the CLUE-S series maps it): of the 1,443 patches of loss, the median has 2 cells and the largest 77, and none contains a 5 × 5 block (500 m a side) that went from all forest to all background. MCD was designed for large clearings seen at fine resolution; at 100 m it cannot confirm a house lot. Yet this loss mattered: the elasticity is 2.76, the interior shrank almost three times as fast as the forest, because every small clearing pierced it. Read the two numbers together: no structural loss and a high elasticity together describe a forest perforated by small clearings.

What the change did to the forest left#

change.fad's stats table has a row per window and one for the average of all (scale 0). The two forests tell different stories:

• Sibuyan lost forest from its edges. The forest left is barely more fragmented at a window of 7 cells (ΔFAD_AV −0.9 points) and only a little more at 27 cells (−3.4), but the density over the whole island fell by 8 points (ΔAVCON), because the forest itself shrank. Of the common forest, between 657 cells (7 cells) and 1,083 (27 cells) moved to another class, every one of them to a more fragmented class (increased_percent = 100): with no forest gained, no cell could become denser.
• Plum Island lost proportionally less forest, but its forest left lost 3.3 to 3.5 points of density at every window: the clearings are everywhere, inside the forest, so they are seen from every forest cell. Here 6 to 14% of the cells that changed class became less fragmented (decreased_percent), next to the 1,270 cells regained.

FOS, cell by cell#

change.fos maps the same change on each cell. It takes two FOS maps, not forest maps: compute fragmentation.fos at both dates with the same settings, then connect their map outputs.

format = "terraflux/1"

[[nodes]]
id = "before"
op = "raster.read"
params = { path = "sibuyan_landuse.tif", band = "1997" }

[[nodes]]
id = "after"
op = "raster.read"
params = { path = "sibuyan_landuse.tif", band = "2011" }

[[nodes]]
id = "fos_before"
op = "fragmentation.fos"
inputs = { map = "before.raster" }
params = { foreground = [1], window = 13 }

[[nodes]]
id = "fos_after"
op = "fragmentation.fos"
inputs = { map = "after.raster" }
params = { foreground = [1], window = 13 }

[[nodes]]
id = "fos_change"
op = "change.fos"
inputs = { before = "fos_before.map", after = "fos_after.map" }

The map's codes need care. On the forest of both dates each cell holds 100 + its connectivity before − its connectivity after (connectivity in %): 100 is no change, above 100 a loss of connectivity, below 100 a gain. Off the common forest: 250 forest gained, 251 forest lost, 252 stable background, 254 missing.

Enlarge

PlayWindow (cells of 250 m)

Enlarge

With a window of 5 cells (1.25 km), 615 of the 4,160 cells of common forest lost connectivity, all within two cells of a clearing; with 27 cells (6.75 km), 2,537 did, a band several kilometres deep behind each front. The halo is the forest that a species sensitive to edges, or to forest cover within its home range, has lost without a tree being cut in it. Choose the window from the species or the process: a home range, a dispersal distance, the reach of edge effects.

How to report it#

For each pair of dates: the forest at both dates, the gain and the loss (simple change), the structural and the spurious change and the elasticity (morphological change), and ΔFAD_AV and ΔAVCON at the windows that matter for your question (FAD's change). Give the maps' source and resolution, the classes counted as forest, and the windows in metres. Show the morphological change map with the simple one.

Settings that change the answer#

• The classes counted as forest (foreground). Plantations, mangroves, woodland: each choice changes every number on this page.
• The resolution. MCD's seeds are 5 × 5 cells, so a coarser map calls more change structural and a finer one more change spurious (compare Sibuyan at 250 m with Plum Island at 100 m). Compare dates only at the same resolution, and say which.
• The windows of FAD and FOS: the slider above. Small windows see the edge, large ones the landscape.
• The order of the dates. before and after are not symmetric: swapping them swaps gain and loss, and a FOS change computed the wrong way round reads every loss of connectivity as a gain.

Morphological change has no setting: its 5 × 5 cores and 3 × 3 closing are GuidosToolbox's.

Common pitfalls#

• GuidosToolbox's legend of FOS change. GuidosToolbox names the groups of the FOS change map "ΔFOS" with the difference taken as before − after: its "ΔFOS +21 to +100" is a loss of 21 to 100 points, coloured dark orange. The legend on this page says it in words; check the sign before writing that connectivity increased.
• Codes that are not GuidosToolbox's. Without foreground, the operators expect 1 background and 2 forest. A land use map with forest as class 1 run without foreground = [1] treats forest as background and coconut as forest.
• Reading spurious as false. Spurious change is unconfirmed, not disproved. In a well-registered pair of maps of the same sensor, much of it is real retreat of the edge.
• Reading structural loss as the area cleared. The closing adds stable cells inside clearings; take the area from simple change.
• Maps of different grids. The operators stop when the two maps differ in size. Resample one onto the other first, with nearest neighbour for classes, and expect a burst of spurious change.
• Missing cells at one date only (a cloud) are missing in the result: compare the missing counts of the stats with your maps.

Going further#

• Part 1: the transition matrix the forest's row and column come from.
• Part 3: map error and the estimation of areas of change from a sample.
• Habitat patterns, part 4: simple and morphological change next to MSPA and the forest's network.
• Fragmentation, part 3: FAD's change at five scales, the change of the landscape mosaic, and GuidosToolbox's reports.
• Operators: change.simple, change.morphological, change.fad, change.fos, fragmentation.fos, change.mosaic.

References#

• Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., & Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148, 42–57. https://doi.org/10.1016/j.rse.2014.02.015
• Riitters, K. H., Wickham, J. D., O'Neill, R. V., Jones, K. B., & Smith, E. R. (2000). Global-scale patterns of forest fragmentation. Conservation Ecology, 4(2), 3. https://doi.org/10.5751/ES-00209-040203
• Riitters, K. H., & Wickham, J. D. (2012). Decline of forest interior conditions in the conterminous United States. Scientific Reports, 2, 653. https://doi.org/10.1038/srep00653
• Riitters, K., Wickham, J., Costanza, J. K., & Vogt, P. (2016). A global evaluation of forest interior area dynamics using tree cover data from 2000 to 2012. Landscape Ecology, 31(1), 137–148. https://doi.org/10.1007/s10980-015-0270-9
• Seebach, L., Strobl, P., Vogt, P., Mehl, W., & San-Miguel-Ayanz, J. (2013). Enhancing post-classification change detection through morphological post-processing – a sensitivity analysis. International Journal of Remote Sensing, 34(20), 7145–7162. https://doi.org/10.1080/01431161.2013.815382
• Townshend, J. R. G., Justice, C. O., Gurney, C., & McManus, J. (1992). The impact of misregistration on change detection. IEEE Transactions on Geoscience and Remote Sensing, 30(5), 1054–1060. https://doi.org/10.1109/36.175340
• Vogt, P., & Riitters, K. (2017). GuidosToolbox: universal digital image object analysis. European Journal of Remote Sensing, 50(1), 352–361. https://doi.org/10.1080/22797254.2017.1330650