Update typology markdown to be consisten with report

Dockerfile

@@ -1,14 +1,14 @@
 FROM ghcr.io/osgeo/gdal:ubuntu-small-3.13.1 AS base
 
 RUN apt-get update && apt-get install -y \
-    python3-pip python3-venv git curl build-essential pkg-config \
-    && apt-get clean && rm -rf /var/lib/{apt,dpkg,cache,log}
+    python3-dev python3-pip python3-venv git curl build-essential pkg-config \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 ENV CARGO_HOME="/usr/local/cargo" RUSTUP_HOME="/usr/local/rustup"
 ENV PATH="$CARGO_HOME/bin:$PATH"
 RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain stable --profile minimal
 
-COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /usr/local/bin/
+COPY --from=ghcr.io/astral-sh/uv:0.10.4 /uv /uvx /usr/local/bin/
 ENV UV_PROJECT_ENVIRONMENT=/code/.venv UV_LINK_MODE=copy
 WORKDIR /code
 

ldn/typology_mapping.yaml

@@ -65,15 +65,15 @@ cci_lc_map:
     82: 1
     90: 1
     100: 1
-    # Grass and Herbaceous / Moss/Lichen -> Grassland
+    # Grass and Herbaceous -> Grassland
     110: 2
     130: 2
-    140: 2
-    # Sparse vegetation -> Grassland
-    150: 2
-    151: 2
-    152: 2
-    153: 2
+    # Moss/Lichen / Sparse vegetation -> other land
+    140: 7
+    150: 7
+    151: 7
+    152: 7
+    153: 7
     # Wetland
     160: 4
     170: 4
@@ -87,7 +87,6 @@ cci_lc_map:
     220: 7
     # Water
     210: 6
-# TODO: CCI needs to map more classes to Other (7). Right now it has no Other pixels in Singapore.
 
 world_cover_map:
     # No data
@@ -127,9 +126,10 @@ io_map:
     5: 3
     # Built-up
     7: 5
-    # Bare/Sparse / Snow/Ice / Cloud -> other land
+    # Bare/Sparse / Snow/Ice -> other land
     8: 7
     9: 7
-    10: 7
+    # Clouds -> no data
+    10: 255
     # Rangeland
     11: 2

typology/typology.md

@@ -1,28 +1,73 @@
-# Preliminary typology proposed for SIDS
-
-The design of the LULC typology should balance model complexity, the feasibility of mapping using satellite data, and the need to capture key classes relevant to SIDS.
-
-This preliminary typology is developed based on a review of existing land-cover products and the class definitions adopted by the UNCCD.
-Level-1 is chosen to align with the UNCCD, for simplicity and to facilitate training data collection.
-
-| Level-1 Class| Proposed Definition | Rationale |
-|---|---|---|
-| Tree cover | Areas dominated by woody vegetation (trees or shrubs), typically forming a continuous or semi-continuous canopy. Includes natural forests, plantations, and shrublands where woody vegetation is the primary cover. | Broadly consistent with the IPCC forest land class but does not enforce specific thresholds (e.g., canopy cover, tree height, minimum area), which require structural information not available from Landsat. Aligns with ESA WorldCover tree cover class by focusing on dominance of woody vegetation, including shrublands. This improves feasibility for satellite-based mapping while maintaining relevance for UNCCD LDN indicators. |
-| Grassland	| Areas dominated by herbaceous vegetation (grasses or forbs), with no or sparse woody cover that does not form a continuous canopy. Includes natural grasslands, rangelands, and pastures.	| Consistent with IPCC grassland definition but operationalized using vegetation dominance rather than land use. Aligns with WorldCover herbaceous vegetation class. The distinction from tree cover is based on absence of continuous woody canopy, which is detectable using spectral-temporal metrics. | 
-| Cropland | Areas used for cultivation of crops, typically showing seasonal vegetation dynamics associated with planting and harvesting cycles. Includes annual crops and herbaceous cropping systems. | Aligned with IPCC cropland but excludes agroforestry systems where woody vegetation is the dominant cover, which are classified under Tree cover. The definition is adapted to rely on seasonal vegetation dynamics rather than land management information, which is not directly observable. Strongly consistent with ESA WorldCover cropland class. |
-| Built-up | Areas dominated by artificial surfaces such as buildings, roads, and infrastructure. May include small patches of vegetation within built environments, but excludes large urban green areas that are predominantly vegetated. | Consistent with IPCC settlements but simplified for remote sensing detection of impervious surfaces. Closely aligned with WorldCover built-up class. The exclusion of large urban green areas ensures consistency with vegetation-based classes and improves separability in Landsat data. |
-| Water | Includes areas covered by surface water for most of the year, such as lakes, rivers and streams, artificial reservoirs, coastal lagoons, and estuaries. | Consistent with UNCCD water bodies definition and WorldCover water class. |
-| Wetland	| Areas where vegetation (woody or herbaceous) is regularly or permanently inundated or saturated with water. Includes mangroves, marshes, and swamps, but excludes open water bodies.	| Aligned with IPCC wetlands (excluding open water) and UNCCD definitions. Harmonized with WorldCover wetland and mangrove classes by emphasizing vegetation under hydrological influence. |
-| Other | Areas with little or no vegetation cover, including bare soil, sand, exposed rock, and ice. | Aligned with IPCC “other land” but narrowed to focus on non-vegetated natural surfaces detectable via remote sensing. |
-
-Level-2 classes are considered aspirational at this stage. The feasibility of mapping Level-2 classes will be assessed during the development phase, guided by stakeholder engagement, the relevance of classes to key degradation processes, and further analytical exploration.
-
-| Level-1 classes| Level-2 sub-classes | Rationale and Development Considerations |
-|---|---|---|
-| Tree Cover | Closed forest, Open forest | Forest classes can be subdivided by canopy cover (open vs. closed), phenology (evergreen vs. deciduous), and leaf type (broadleaved vs. needleleaved). Different LULC products have adopted different combinations of these subdivisions. Across SIDS, biome diversity is relatively limited, and temporal profiles are often less distinct due to persistent cloud cover and sparse time-series observations. Since we expect to rely more heavily on spectral characteristics than on multi-temporal signatures, crown density (open vs. closed forest) is the most feasible and robust Level-2 subdivision for SIDS. In the Pacific, subclasses of native evergreen forest and secondary forest are proposed. This separation is not always detectable using moderate-resolution remote sensing.  |
-| Grassland | Shrubland, Grassland | Most global LULC products treat shrubland and grassland as top-level categories, but in practice, these classes are frequently confused, especially in heterogeneous dryland or coastal environments common across SIDS. Despite the classification challenges, distinguishing these classes remains potentially valuable for ecosystem management, erosion monitoring, and land degradation reporting. |
-| Cropland | Rainfed, Irrigated| The rainfed/irrigated subdivision used by products such as GLC_FCS30D relies heavily on multi-temporal patterns. For many SIDS, persistent cloud cover, small field sizes, and limited temporal depth make it difficult to separate irrigation regimes reliably. Given the challenges of identifying cropland as a broad class, this subclass distinction will only be attempted if irrigation status is determined as particularly relevant for SIDS. This separation has not been proposed in the Pacific. Instead, subclasses that are more relevant for LDN and may be detectable from remote sensing include perennial tree crops and agroforestry. However, these systems can be easily confused with native tree cover. |
-| Built-up | - | |
-| Water | - | |
-| Wetland | Herbaceous wetland, Mangrove | Products such as WorldCover separate herbaceous wetlands and mangroves into distinct classes. For SIDS, coastal ecosystems, especially mangroves, are of exceptionally high ecological and socio-economic importance, providing coastal protection, carbon storage, and nursery grounds for marine species. Given this importance and the relatively clear spectral separability of mangroves, maintaining mangroves as a dedicated Level-2 subclass is both feasible and highly beneficial for monitoring and conservation objectives. |
-| Other | - | |
+# Land Use Land Cover Typology for SIDS
+
+This document defines the Land Use/Land Cover (LULC) typology for a 30 m LULC product for Small Island Developing States (SIDS), developed to support the Land Degradation Neutrality (LDN) initiative (CI GEF project 11834). Annual maps will be generated for 2000-2025 from Landsat geometric median and median absolute deviation composites.
+
+The typology is a simple hierarchy: a consistent set of Level-1 classes for LDN reporting, plus a small number of Level-2 subclasses where they are feasible to map and locally relevant. Level-2 classes aggregate to Level-1.
+
+## Design requirements
+
+The typology should:
+
+- support calculation of the LDN indicators
+- enable assessment of the spatial extent and key drivers of land degradation
+- capture country-specific land degradation processes, as emphasised in UNCCD guidance
+- be feasible to map using remotely sensed data
+- be mutually exclusive and collectively exhaustive, so every valid pixel gets one and only one class
+- balance classification complexity with the suitability and availability of input data
+- be compatible, where practicable, with classification systems already used by relevant countries and existing data-collection efforts
+
+Level-1 follows the UNCCD default seven-class typology (adapted from the IPCC land-use categories), so data can serve multiple reporting purposes and IPCC land-use change factors can be applied to soil organic carbon estimates. Definitions are adapted to be practical for satellite-based mapping.
+
+## Level-1 classes
+
+| Level-1 Class | Proposed Definition | Rationale |
+| ------------- | ------------------- | --------- |
+| Tree cover | Areas dominated by trees, typically forming a continuous or semi-continuous canopy. Includes natural forests, plantations and tree crops where tree cover is dominant, but excludes shrub-dominated vegetation and mangroves. | Broadly consistent with the IPCC Forest Land category but does not enforce thresholds for canopy cover, tree height or minimum area, as these cannot all be determined consistently from Landsat alone. Consistent with ESA WorldCover in distinguishing tree cover from shrubland and mangroves. |
+| Grassland | Areas dominated by herbaceous vegetation or shrubs that do not meet the Tree cover definition. Includes natural grasslands, rangelands, pastures and shrubland. | Broadly consistent with the IPCC and UNCCD Grassland category, which includes shrub-dominated land not classified as Forest Land, but adapted to focus on vegetation dominance rather than land use. Corresponds broadly to the combined ESA WorldCover Grassland and Shrubland classes. |
+| Cropland | Areas used for cultivation of crops, typically showing seasonal vegetation dynamics associated with planting and harvesting cycles. Includes annual crops and herbaceous cropping systems. | Aligned with IPCC cropland but excludes agroforestry systems where woody vegetation is the dominant cover, which are classified as Tree cover. Relies on seasonal vegetation dynamics rather than land management information, which is not directly observable. Broadly consistent with the ESA WorldCover cropland class. |
+| Built-up | Areas dominated by artificial surfaces such as buildings, roads and infrastructure. May include small patches of vegetation within built environments, but excludes large urban green areas that are predominantly vegetated. | Consistent with IPCC settlements but simplified for remote sensing detection of impervious surfaces. Closely aligned with the WorldCover built-up class. Excluding large urban green areas keeps it consistent with the vegetation-based classes and improves separability in Landsat data. |
+| Water | Areas covered by surface water for most of the year, such as lakes, rivers and streams, artificial reservoirs, coastal lagoons and estuaries. | Consistent with the UNCCD water bodies definition and the WorldCover water class. |
+| Wetland | Areas where vegetation (woody or herbaceous) is regularly or permanently inundated or saturated with water. Includes mangroves, marshes and swamps, but excludes open water bodies. | Broadly consistent with IPCC wetlands (excluding open water) and UNCCD definitions. Harmonised with the WorldCover wetland and mangrove classes by emphasising vegetation under hydrological influence. |
+| Other | Areas with little or no vegetation cover, including bare soil, sand, exposed rock and ice. | Aligned with IPCC "other land" but narrowed to non-vegetated natural surfaces detectable via remote sensing. |
+
+## Level-2 classes
+
+A minimal hierarchy is proposed, with Shrubland and Mangrove as the only Level-2 classes. Both are particularly relevant to land degradation assessment in Pacific SIDS and are represented in existing Pacific field-data collection.
+
+| Level-1 Class | Level-2 Class | Proposed Definition |
+| ------------- | ------------- | ------------------- |
+| Grassland | Shrubland | Areas dominated by shrubs or other low woody vegetation, where trees are not the dominant cover. |
+| Wetland | Mangrove | Intertidal wetlands dominated by mangrove vegetation. |
+
+Shrubland is mapped as a distinct Level-2 class, consistent with ESA WorldCover, and aggregated with herbaceous Grassland to derive Level-1 Grassland. Mangrove is mapped separately and aggregated with other vegetated wetlands to derive Level-1 Wetland.
+
+**Status: provisional.** Inclusion of Shrubland and Mangrove in the final product depends on stakeholder confirmation of their decision relevance and on pilot testing showing they can be interpreted and mapped consistently. Additional subclasses, including tree crops, may be considered later where there is a demonstrated decision need and adequate reference data.
+
+## Implementation rules
+
+These rules guide both map production and validation-data collection.
+
+1. **Dominant observable cover.** Label each pixel by its dominant observable land cover during the reference period. Where multiple covers occur, assign the class with the largest estimated cover fraction. Dominance does not require more than 50% cover. This is needed because the product has no mixed-pixel labels.
+2. **Hydrologically influenced areas.** Open water is Water. Areas dominated by vegetation under regular or persistent hydrological influence are Wetland. Mangroves are always Wetland.
+3. **Cultivated vegetation.** Annual and herbaceous cropping systems are Cropland. Tree crops and agroforestry are classified by dominant observable cover, so areas dominated by trees are Tree cover.
+4. **Modified non-vegetated surfaces.** Areas dominated by buildings, roads or other constructed surfaces are Built-up. Extraction areas, stockpiles, waste sites and cleared ground are classified by dominant observable cover, generally Other unless constructed surfaces dominate.
+5. **Temporary conditions.** Temporary flooding, burning, exposed soil or vegetation loss should not automatically determine the annual class. Classification should reflect dominant or persistent cover during the reference period, using multi-date observations where possible.
+6. **Uncertain reference labels.** Apply the same rules when collecting validation data. Interpreters should record label confidence and, where useful, secondary cover or estimated cover fractions. Samples that cannot be labelled reliably are flagged for review or excluded per the validation protocol.
+
+These rules will be tested through mapping and validation. Recurring ambiguities may require more detailed class-specific decision rules or targeted use of ancillary and higher-resolution reference data.
+
+## Training labels and known classification issues
+
+Initial training data will come from agreement among ESA CCI Land Cover, ESA WorldCover and Impact Observatory (IO) LULC, reclassified as below.
+
+| Level-1 Class | ESA WorldCover classes | IO LULC classes | ESA CCI Land Cover classes |
+| ------------- | ---------------------- | --------------- | -------------------------- |
+| Tree cover | Tree cover | Trees | Tree cover (all leaf types and densities), Mosaic tree and shrub |
+| Grassland | Shrubland, Grassland | Rangeland | Mosaic natural vegetation / cropland, Mosaic herbaceous / tree and shrub, Shrubland, Grassland |
+| Cropland | Cropland | Crops | Cropland (rainfed and irrigated), Mosaic cropland / natural vegetation |
+| Built-up | Built-up | Built Area | Urban areas |
+| Other | Bare/sparse vegetation, Snow and Ice, Moss and lichen | Bare Ground, Snow/Ice | Lichens and mosses, Sparse vegetation, Bare areas, Permanent snow and ice |
+| Water | Permanent water bodies | Water | Water bodies |
+| Wetland | Herbaceous wetland, Mangroves | Flooded Veg. | Tree cover flooded (fresh and saline), Shrub or herbaceous cover flooded |
+
+These products differ in resolution, reference year, class definitions, thresholds and methods, so the class definitions cannot be applied perfectly to every derived label, and some ambiguity will propagate into the maps. Validation data must apply the project definitions independently of the training labels, so the accuracy assessment captures errors from ambiguous training labels rather than reproducing them.