Preliminary LULC (v0-0-9) assessment

[fangfy]

List of issues identified across SPC countries in LULC v0-0-9 output.

New Caledonia

Positives:

• General patterns of tree cover, grassland, wetland and water look good and reasonably stable between 2000 and 2025.
• Major roads correctly identified as built-up.
• Crop lands near settlements, sometimes quite scattered, are picked up.

1. False positive built-up areas, confusion with bright bare surfaces
   E.g. this rocky island

Doesn't happen when the bare surface is not as bright/white.

Possible cause/solution: underrepresented "other" class in training? Do existing LC maps classify correctly over these areas? Ensure similar bare surfaces are included in training.

2. False positive cropland, confusion with grassland/shrubs
   e.g. this area far from any settlements

Possible cause/solution: Separating cropland and grassland is a known challenge. It might be that misclassification occurs when the seasonal change signature is more pronounced than expected. Need to examine the training data for these two classes to see whether the features show clear separability.

Local knowledge on where cropping and grazing happen would be extremely valuable.

[fangfy]

Tonga

Excellent classification for settlements. False positive built-up areas over bright beaches.
Great identification of cropland.

Possible confusion between cropland and tree cover, possibly due to mixed farming/small fields and/or tree crops.
e.g. around Matahau, Tonga

Worth a discussion of what the best classification system should be used here. Should the whole area be cropland? Or is this mixed classification OK. The classification for land cover (tree or grass) is actually quite accurate but it doesn't consider land use.

[fangfy]

Samoa

Grassland or cropland?
Near Apia, Samoa

Tree cover and clearing along the roads are generally picked up well (resolution allows). Hard to tell whether the grass areas are cultivated. There are also likely tree crops (e.g. coconut) that are correctly classified as tree cover.

[fangfy]

French Polynesia

Terrain shadow misclassified as water
Mountain region

Some waterbodies are actually correctly identified along the valleys; however, anything on steep slopes is likely false. Could probably be filtered out by just a slope mask, if the DEM is good enough for these areas.

[fangfy]

PNG

Very challenging with clouds.

Mangroves are generally picked up well
along the south coast of PNG
Global Mangrove Watch

Cropland and built-up classification are likely noise. Inland wetland and grassland look reasonable

[willjnz]

Non-Pacific LULC for 2000 and 2025 now available. App now allows comparing data across years: https://mmufb4pjqf.execute-api.us-west-2.amazonaws.com/?pos=20.1539%2C282.0081%2C7&year=2025&basemap=hybrid&layers=class&op=class%3A1&left=class%3A2000%3A1&right=class%3A2025%3A1