Methods if You Are Making And Using a Stratification Map

The plan discussed in the workshop event was to use the land cover maps created using Random Forest, comparing them to make LC Change Maps to be used as the stratification maps for sample distribution. This could be done in GEE or QGIS, so we will discuss both options.

Overview

Post-classification comparison is a straightforward approach for generating change maps from two land cover (LC) datasets (t1 and t2). Remember, these are the labels associated wih your class numbers for the LC maps.

Numeric code	Class name
0	nodata
1	forest_80
2	forest_30_80
3	forest_30
4	mangroves
5	settlements
7	water
8	grassland
9	shrub
10	baresoil
11	sand

To make the change maps, you compare the class value of each pixel at time 1 against its value at time 2 and then assign a change label. This can be done with either:

• Full-detail transitions: every possible class-to-class change (e.g., forest_80 → settlements).
• Simplified schemes:
  • binary (change / no_change), or
  • thematic categories (e.g., no_change, forest_loss, mangrove_loss, other_change).

Key considerations for your change maps

Input maps

• Same projection, resolution, and extent.
• Same class scheme (you already have consistent codes: 1–11).
• Make sure it’s clear which map is t1 (earlier) and which is t2 (later).

Change legend design

• Full detail: every possible transition (e.g., forest_80 → settlements, mangroves → water). This is rich but can produce many classes.
• Simplified: e.g.,
  • no_change
  • forest_loss (any forest/mangrove class → non-forest)
  • mangrove_loss (mangroves → non-mangroves)
  • other_change (grassland → shrub, shrub → forest, etc.)
• Or the most minimal: 2 classes – no_change, change.

---

Option 1 - GEE Method

General Approach

Use the two LC rasters (lc_t1, lc_t2) to calculate transition codes or classify changes directly using logical expressions.

A. Detailed Transition Map (all change types)

Each transition can be uniquely encoded:

var transition = lc_t1.multiply(100).add(lc_t2);
// e.g., forest_80 (1) → settlements (5) becomes 105

You can use this map as-is or reclassify it later into thematic change groups.

B. Simplified Change Classes

Define forest and mangrove groups:

var forestClasses = [1,2,3]; // forest strata
var isForest_t1 = lc_t1.remap(forestClasses, [1,1,1], 0); var isForest_t2 = lc_t2.remap(forestClasses, [1,1,1], 0);
var isMangrove_t1 = lc_t1.eq(4); var isMangrove_t2 = lc_t2.eq(4);
var noChange = lc_t1.eq(lc_t2);

Build a simplified 4-class scheme:

var changeClass =
noChange.multiply(1) // 1 = no_change .where(isForest_t1.eq(1).and(isForest_t2.eq(0)), 2) // forest_loss .where(isMangrove_t1.eq(1).and(isMangrove_t2.eq(0)), 3) // mangrove_loss .where(noChange.not()

.and(isForest_t1.eq(0).or(isForest_t2.eq(0))) .and(isMangrove_t1.eq(0).and(isMangrove_t2.eq(0))), 4); // other_change

C. Binary Change / No-Change Map

Updated per your request: 0 = no_change, 1 = change.

var changeBinary = lc_t1.neq(lc_t2).rename('change'); // changeBinary: 1 = change, 0 = no_change

D. Basic Steps in GEE

1. Load LC maps (lc_t1, lc_t2), ensure same projection/resolution.
2. Compute either:
   • a full transition image
   • simplified change classes
   • or a binary chang/no-change map
3. Visualize and export.

A full demo script of this process is in the GEE repository (CreateChangeMapForStratification). Quick link.

---

Option 2 - QGIS Method

Ensure LC rasters from t1 and t2 are aligned. In QGIS, you’ll typically use Raster Calculator and/or “Raster → Raster Calculator” and the “Reclassify by table” or “Raster reclassify” tools (from Processing). Use Raster Calculator with logical expressions.

A. Binary Change / No-Change

In Raster Calculator:

("LC_t1@1" != "LC_t2@1")

Save as change_binary.tif

B. Detailed Transition Codes

In Raster Calculator, if using detailed transition codes:

("LC_t1@1" * 100) + "LC_t2@1"

The first map label is seen in the 10s place, while the second map is the 1s digit. This produces codes such as:

• 105 = 1 to 5 transition
• 403 = 4 to 3 transition

These can be reclassified with a lookup table to thematic groups. Use “Reclassify by table” or “Raster reclassify” tools (from Processing). Use a reclassification table (CSV or manually in the tool) if you want:

• to group transitions (e.g., all forest→non-forest as forest_loss), or
• to directly relabel specific transitions.

C. Simplified Change Classes

Decide which LC codes are forest/mangroves:

• Forest = {1,2,3}
• Mangroves = {4}
• Non-forest = {5, 7, 8, 9, 10, 11}

Use nested if() expressions in Raster Calculator:

if( "LC_t1@1" = "LC_t2@1", 1,     -- no_change
    if( ("LC_t1@1" in (1,2,3)) AND NOT ("LC_t2@1" in (1,2,3)),
        2,                       -- forest_loss
        if( ("LC_t1@1" = 4) AND ("LC_t2@1" != 4),
            3,                   -- mangrove_loss
            4                    -- other_change
        )
    )
)

Save as change_classes.tif and apply a custom legend.

• 1 = no_change
• 2 = forest_loss
• 3 = mangrove_loss
• 4 = other_change

D. Basic Steps in QGIS

1. Load rasters.
2. Use Raster Calculator or “Reclassify by table” to compute:
   • binary change
   • transition codes
   • or simplified thematic classes
3. Style and export.

Using AREA2 for Sample Distribution

You will most likely be using Stratified Random Sampling (using a map to inform the design, a minimum number of observations randomly placed in each category), so that is assumed here.

• If you decide to use random or systematic sampling you can do that in QGIS or directly in CEO. All you would need is a shapefile for your AOI.

Stratified random sampling key benefits:

• Updates map-based areas to increase precision (reduces uncertainty)
• Helps increase chance of having plots in rare classes
• A map is not necessary for selecting sample locations but it makes the representative sampling process far more efficient when you are looking at small areas of change, reducing the total number of points you need to collect to meet your desired level of uncertainty

Now that you have a stratification map, you can use AREA2 to distribute your samples. Refer to the full directions on AREA2 on the 3. Sampling Design / Sampling Design using AREA2 page. Remember, we have a local version of AREA2 set up in your GEE repository for this workshop.

We reviewed how to use AREA2 in a follow-up workshop to the main workshop event. The recording of that walk-through is here. You can stop watching at 16:00, when the topic switches to uploading the samples to CEO. For the area estimation of land cover change you will be using a different map asset than was used in the video.

The Google Sheets template we used in the follow-up workshop to calculate the number of samples per strata (proportional with a minimum sample size) is here. You can make a copy of the template in order to make your own edits. As seen in the demo video and previous sections, a key point to remember is that you must have 30 samples minimum per strata, but 50 or more is better in order for the statistical equations for area estimation to be valid. Therefore, the more strata in your map, the more points you will need to collect. But having more detailed strata can be beneficial depending on your goals. The total number of samples you collect can be determined based on a target minimim viable uncertainty, as is done in this spreadhseet, or simply be based on available resources (the number of interpreters and the time available for data collection). As a rule of thumb, 1,000 points for your Liberia AOI should probably be the minimum for decent results.

The major objective of the sample is to provide information that is representative of the population. If the sample is representative of the population, then the sample statistics are good and valid estimates of the population parameters.

However, sampling can be expensive, so you must balance sampling validity with cost when considering total sample size and stratification. Consider the time it will take to collect the samples plus time necessary for QA/QC reviews.