Texte exact envoyé au modèle pour chaque variante.
hierarchical — Hierarchicalhaiku_depth — Haiku + depthhierarchical_en_no_comment — Hierarchical (no comment)gt_aligned — GT alignedhierarchical — HierarchicalYou are classifying ground-level street-view photos from rural Benin (Parakou landscapes).
The photo is taken from a road. IGNORE the road strip at the bottom unless the entire image is road or bare soil.
Classify LAND USE of the main plot, not the tallest object.
Scattered trees are normal in this landscape. Trees in FRONT of the plot, trees BEHIND the plot, and roadside trees must not steal the label from an identifiable annual crop.
These labels will train satellite land-cover models. If you cannot see the plot, or two land uses are co-dominant, return Unclassifiable. Do not guess. Unclassifiable is not a land-cover class and must not be used as satellite ground truth.
## Step A — coarse class (mandatory)
Exactly one of:
- built_up
- bare_road
- water
- annual_crop
- tree_crop
- natural_vegetation
- unclassifiable
## Step B — fine class (mandatory)
class_name_en MUST be copied EXACTLY from the list under the coarse class you chose. Never invent a name. Never use "Wooded fallow", "Tree savanna", "Shrubby fallow", "Other Crop", "Other Orchard", "Other Class", or "Agroforestry".
### built_up
- Built-up — houses, shops, walls, paved compounds, village fabric.
### bare_road
- Bare road — dirt/laterite road or bare earth dominates the frame.
### water
- Water — open water is the subject.
### annual_crop (use this if a cultivated herbaceous/root field is identifiable on the ground)
- Maize — tall broad-leaf cereal, often in rows; cobs or tassels if mature; looks like large grass, not a grain-head panicle.
- Sorghum/Millet — tall cereal with a compact or broom-like GRAIN HEAD / panicle at the tip; narrower leaves than maize.
- Soybean — low-to-medium dense uniform field of bushy legumes, trifoliate leaves, no mounds, no panicles, no maize stalks. Often looks like a continuous green carpet ~0.4–1 m. If it is clearly a managed herbaceous field but you cannot tell soybean vs mixed legumes, use Other annual herbaceous crop instead of natural vegetation.
- Yam — regular soil MOUNDS / ridges (buttes), often vines or stakes.
- Cassava — woody shrubs 1–3 m, palmate (hand-shaped) leaves, often irregular spacing in a field.
- Rice — very dense grass in a wet/lowland plot; standing water or bunds.
- Cotton — bushy shrubs; white bolls only if mature.
- Cowpea — climbing/spreading legume, often intercropped; use only if clearly cowpea.
- Groundnut — very low groundnut carpet; rare in this set — prefer Other annual herbaceous crop unless sure.
- Other annual herbaceous crop — cultivated herbaceous field that is not clearly maize/sorghum/soybean/yam/cassava/rice.
### tree_crop (planted trees ARE the land use: monospecific orchard/plantation)
- Cashew — medium trees, rounded crown, compound leaves, often yellowish; look for cashew apples/nuts; spacing often regular but young plots can look scrubby. Do not require fruit.
- Mango — large dark evergreen dome, much bigger than cashew.
- Teak — tall straight trunks, very large broad opposite leaves, often in rows; dry-season yellowish large leaves.
- Gmelina arborea — gmelina plantation (similar to teak but typically smaller/less columnar). Rare; prefer Teak if unsure between the two plantations.
- Other orchard — other planted orchard (citrus, oil palm, mixed fruit) that is not cashew/mango/teak.
### natural_vegetation (no identifiable annual crop, and not a planted orchard)
- Forest — closed or nearly closed tree canopy; forest interior or dense woodland; little ground visible.
- Herbaceous savanna — grass-dominated, few or no trees/shrubs, not a crop.
- Shrubby vegetation — shrubland, open woodland, tree savanna, or parkland with scattered trees over grass/weeds. NOT a planted orchard and NOT closed forest. There is no Agroforestry class: parkland without an identifiable annual crop is Shrubby vegetation.
### unclassifiable
- Unclassifiable — do not assign a land-use class. Set usable_for_satellite_gt to false and pick exactly one reject_reason:
- unusable_image — blur, darkness, glare, only sky/road, vehicle, people filling the frame
- ground_not_visible — trees/vegetation in front hide the plot; you cannot see the ground layer
- mixed_plots — two land uses are co-dominant (field boundary, mosaic); no single plot dominates
- unknown_land_use — image is usable but you cannot tell the land use
## Decision rules (in order)
0. If the photo is unusable, the ground of the plot is not visible, or two land uses are co-dominant → unclassifiable. Do not pick a compromise class.
1. If buildings dominate → Built-up. If the photo is mostly road/bare laterite → Bare road.
2. If an annual crop is identifiable on the ground (rows, mounds, uniform legumes, cereal heads, cassava shrubs, rice, etc.) → annual_crop. This holds even if trees stand in FRONT of the crop, BEHIND it, or along the road. Do not switch to natural_vegetation, tree_crop, or Unclassifiable because trees are visible if the crop on the ground is still identifiable.
3. If trees are planted as the land use (same species, regular spacing, plantation/orchard form) → tree_crop, even if the understory is weedy or grassy.
4. Closed canopy, no field → Forest. Open grass, almost no woody plants → Herbaceous savanna. Mixed shrubs/trees, parkland, or uncultivated open woodland → Shrubby vegetation.
5. NEVER use natural_vegetation as the default because "there are trees". Scattered trees + a visible annual crop is still annual_crop. Trees in front that fully hide the ground → Unclassifiable (ground_not_visible), not Shrubby vegetation.
6. Confidence: ≥0.85 only with organ-level cues (panicles, cobs, mounds, cashew apples, teak leaves, palmate cassava). Species guess without those cues: 0.4–0.7. Coarse class can be more confident than the fine class. usable_for_satellite_gt is true only for a real land-use class (not Unclassifiable).
JSON only (no markdown):
{"level1":"<coarse>","level1_confidence":<0-1>,"class_name_en":"<exact fine label>","confidence":<0-1>,"cues":["<cue>"],"reasoning":"<1-2 sentences>","usable_for_satellite_gt":<true|false>,"reject_reason":null|"unusable_image"|"ground_not_visible"|"mixed_plots"|"unknown_land_use"}
haiku_depth — Haiku + depthYou are classifying ground-level street-view photos from rural Benin (Parakou landscapes).
The photo is taken from a road. IGNORE the road strip at the bottom unless the entire image is road or bare soil.
Classify LAND USE of the main plot, not the tallest object.
Scattered trees are normal in this landscape. Trees in FRONT of the plot, trees BEHIND the plot, and roadside trees must not steal the label from an identifiable annual crop.
These labels will train satellite land-cover models. If you cannot see the plot, or two land uses are co-dominant, return Unclassifiable. Do not guess. Unclassifiable is not a land-cover class and must not be used as satellite ground truth.
## Step A — coarse class (mandatory)
Exactly one of:
- built_up
- bare_road
- water
- annual_crop
- tree_crop
- natural_vegetation
- unclassifiable
## Step B — fine class (mandatory)
class_name_en MUST be copied EXACTLY from the list under the coarse class you chose. Never invent a name. Never use "Wooded fallow", "Tree savanna", "Shrubby fallow", "Other Crop", "Other Orchard", "Other Class", or "Agroforestry".
### built_up
- Built-up — houses, shops, walls, paved compounds, village fabric.
### bare_road
- Bare road — dirt/laterite road or bare earth dominates the frame.
### water
- Water — open water is the subject.
### annual_crop (use this if a cultivated herbaceous/root field is identifiable on the ground)
- Maize — tall broad-leaf cereal, often in rows; cobs or tassels if mature; looks like large grass, not a grain-head panicle.
- Sorghum/Millet — tall cereal with a compact or broom-like GRAIN HEAD / panicle at the tip; narrower leaves than maize.
- Soybean — low-to-medium dense uniform field of bushy legumes, trifoliate leaves, no mounds, no panicles, no maize stalks. Often looks like a continuous green carpet ~0.4–1 m. If it is clearly a managed herbaceous field but you cannot tell soybean vs mixed legumes, use Other annual herbaceous crop instead of natural vegetation.
- Yam — regular soil MOUNDS / ridges (buttes), often vines or stakes.
- Cassava — woody shrubs 1–3 m, palmate (hand-shaped) leaves, often irregular spacing in a field.
- Rice — very dense grass in a wet/lowland plot; standing water or bunds.
- Cotton — bushy shrubs; white bolls only if mature.
- Cowpea — climbing/spreading legume, often intercropped; use only if clearly cowpea.
- Groundnut — very low groundnut carpet; rare in this set — prefer Other annual herbaceous crop unless sure.
- Other annual herbaceous crop — cultivated herbaceous field that is not clearly maize/sorghum/soybean/yam/cassava/rice.
### tree_crop (planted trees ARE the land use: monospecific orchard/plantation)
- Cashew — medium trees, rounded crown, compound leaves, often yellowish; look for cashew apples/nuts; spacing often regular but young plots can look scrubby. Do not require fruit.
- Mango — large dark evergreen dome, much bigger than cashew.
- Teak — tall straight trunks, very large broad opposite leaves, often in rows; dry-season yellowish large leaves.
- Gmelina arborea — gmelina plantation (similar to teak but typically smaller/less columnar). Rare; prefer Teak if unsure between the two plantations.
- Other orchard — other planted orchard (citrus, oil palm, mixed fruit) that is not cashew/mango/teak.
### natural_vegetation (no identifiable annual crop, and not a planted orchard)
- Forest — closed or nearly closed tree canopy; forest interior or dense woodland; little ground visible.
- Herbaceous savanna — grass-dominated, few or no trees/shrubs, not a crop.
- Shrubby vegetation — shrubland, open woodland, tree savanna, or parkland with scattered trees over grass/weeds. NOT a planted orchard and NOT closed forest. There is no Agroforestry class: parkland without an identifiable annual crop is Shrubby vegetation.
### unclassifiable
- Unclassifiable — do not assign a land-use class. Set usable_for_satellite_gt to false and pick exactly one reject_reason:
- unusable_image — blur, darkness, glare, only sky/road, vehicle, people filling the frame
- ground_not_visible — trees/vegetation in front hide the plot; you cannot see the ground layer
- mixed_plots — two land uses are co-dominant (field boundary, mosaic); no single plot dominates
- unknown_land_use — image is usable but you cannot tell the land use
## Decision rules (in order)
0. If the photo is unusable, the ground of the plot is not visible, or two land uses are co-dominant → unclassifiable. Do not pick a compromise class.
1. If buildings dominate → Built-up. If the photo is mostly road/bare laterite → Bare road.
2. If an annual crop is identifiable on the ground (rows, mounds, uniform legumes, cereal heads, cassava shrubs, rice, etc.) → annual_crop. This holds even if trees stand in FRONT of the crop, BEHIND it, or along the road. Do not switch to natural_vegetation, tree_crop, or Unclassifiable because trees are visible if the crop on the ground is still identifiable.
3. If trees are planted as the land use (same species, regular spacing, plantation/orchard form) → tree_crop, even if the understory is weedy or grassy.
4. Closed canopy, no field → Forest. Open grass, almost no woody plants → Herbaceous savanna. Mixed shrubs/trees, parkland, or uncultivated open woodland → Shrubby vegetation.
5. NEVER use natural_vegetation as the default because "there are trees". Scattered trees + a visible annual crop is still annual_crop. Trees in front that fully hide the ground → Unclassifiable (ground_not_visible), not Shrubby vegetation.
6. Confidence: ≥0.85 only with organ-level cues (panicles, cobs, mounds, cashew apples, teak leaves, palmate cassava). Species guess without those cues: 0.4–0.7. Coarse class can be more confident than the fine class. usable_for_satellite_gt is true only for a real land-use class (not Unclassifiable).
JSON only (no markdown):
{"level1":"<coarse>","level1_confidence":<0-1>,"class_name_en":"<exact fine label>","confidence":<0-1>,"cues":["<cue>"],"reasoning":"<1-2 sentences>","usable_for_satellite_gt":<true|false>,"reject_reason":null|"unusable_image"|"ground_not_visible"|"mixed_plots"|"unknown_land_use"}
## Extra input — aligned depth map
You are given TWO images of the SAME view:
1. RGB street-view photograph.
2. Metric depth rendered with a turbo colormap: dark purple / blue = NEAR
(foreground), yellow / red = FAR (background). Invalid pixels are black.
Use depth as a geometric cue, not as a second scene:
- Trees or shrubs that are much nearer than the field behind them are FOREGROUND,
not the plot land use.
- A closed nearby canopy (near values filling the frame) often means the plot
ground is not visible → unclassifiable / ground_not_visible.
- Regular far-range tree spacing can support tree_crop; irregular near/far mix
of woody plants over grass supports natural_vegetation.
Do not ignore a clearly identifiable annual crop on the ground just because
trees appear in the RGB image.
hierarchical_en_no_comment — Hierarchical (no comment)You are classifying ground-level street-view photos from rural Benin (Parakou landscapes).
The photo is taken from a road. IGNORE the road strip at the bottom unless the entire image is road or bare soil.
Classify LAND USE of the main plot, not the tallest object.
Trees in FRONT of the plot, trees BEHIND the plot, and roadside trees must not steal the label from an identifiable annual crop.
These labels will train satellite land-cover models. If you cannot see the plot, return unclassifiable. Never use Agroforestry.
## Step A — coarse class (mandatory)
Exactly one of:
- built_up
- bare_road
- water
- annual_crop
- tree_crop
- natural_vegetation
- unclassifiable
## Step B — fine class (mandatory)
class_name_en MUST be copied EXACTLY from the list under the coarse class you chose. Never invent a name. Never use "Wooded fallow", "Tree savanna", "Shrubby fallow", "Other Crop", "Other Orchard", "Other Class", or "Agroforestry".
### built_up
- Built-up
### bare_road
- Bare road
### water
- Water
### annual_crop
- Maize
- Sorghum/Millet
- Soybean
- Yam
- Cassava
- Rice
- Cotton
- Cowpea
- Groundnut
- Other annual herbaceous crop
### tree_crop
- Cashew
- Mango
- Teak
- Gmelina arborea
- Other orchard
### natural_vegetation
- Forest
- Herbaceous savanna
- Shrubby vegetation
### unclassifiable
- Unclassifiable
## Decision rules (in order)
0. Unusable photo, ground of the plot not visible, or two land uses co-dominant → unclassifiable (usable_for_satellite_gt=false; reject_reason=unusable_image|ground_not_visible|mixed_plots|unknown_land_use).
1. If buildings dominate → Built-up. If the photo is mostly road/bare laterite → Bare road.
2. If an annual crop is identifiable on the ground → annual_crop, even if trees stand in FRONT or BEHIND.
3. If trees are planted (spacing, same species, plantation form) → tree_crop, even if the understory is weedy or grassy.
4. Closed canopy, no field → Forest. Open grass → Herbaceous savanna. Mixed shrubs/trees, parkland, uncultivated open woodland → Shrubby vegetation.
5. NEVER use natural_vegetation as the default because "there are trees". Trees in front that fully hide the ground → Unclassifiable, not Shrubby vegetation.
6. Confidence: ≥0.85 only with organ-level cues (panicles, cobs, mounds, cashew apples, teak leaves, palmate cassava). Species guess without those cues: 0.4–0.7. Coarse class can be more confident than the fine class.
JSON only (no markdown):
{"level1":"<coarse>","level1_confidence":<0-1>,"class_name_en":"<exact fine label>","confidence":<0-1>,"cues":["<cue>"],"reasoning":"<1-2 sentences>","usable_for_satellite_gt":<true|false>,"reject_reason":null|"unusable_image"|"ground_not_visible"|"mixed_plots"|"unknown_land_use"}
gt_aligned — GT alignedYou are classifying ground-level street-view photos from rural Benin.
Output class_name_en MUST be exactly one of these labels (copy spelling):
Built-up, Bare road, Water,
Maize, Sorghum/Millet, Soybean, Yam, Cassava, Rice, Cotton, Cowpea, Groundnut, Other annual herbaceous crop,
Cashew, Mango, Teak, Gmelina arborea, Other orchard,
Forest, Herbaceous savanna, Shrubby vegetation,
Unclassifiable
There is no Agroforestry class. Parkland without an identifiable annual crop → Shrubby vegetation.
Visual guide:
- Built-up: buildings, village, walls, metal roofs.
- Bare road: dirt/laterite road or bare soil is the subject.
- Water: water body.
- Maize: tall broad-leaf cereal, rows, cobs/tassels if mature.
- Sorghum/Millet: tall cereal with a grain panicle/head at the top (not maize).
- Soybean: dense low bushy legume field, trifoliate leaves, no mounds, no panicles. If it is a managed green field but species is unclear → Other annual herbaceous crop, NOT savanna.
- Yam: soil mounds/ridges, vines/stakes.
- Cassava: 1–3 m shrubs, palmate leaves.
- Rice: wet dense grass, paddies/lowland.
- Cotton: bushy; bolls if mature. Rare.
- Cowpea / Groundnut: legumes; rare — prefer Other annual herbaceous crop if unsure.
- Other annual herbaceous crop: other or mixed annual crop.
- Cashew: cashew orchard (spacing, compound leaves, cashew apples). Do not require fruit. Do not dump cashew into shrubland.
- Mango: large dark mango crowns.
- Teak: teak plantation, huge leaves, straight trunks, rows.
- Gmelina arborea: gmelina plantation. Rare.
- Other orchard: other orchard.
- Forest: closed tree canopy. Dense woodland is Forest, not shrubland.
- Herbaceous savanna: grass savanna, few trees, not a crop.
- Shrubby vegetation: shrubs / open woodland / parkland, not orchard, not forest.
- Unclassifiable: unusable photo, ground of the plot not visible, mixed plots, or unknown land use. Not satellite ground truth.
Rules:
1. Classify plot land use; ignore the foreground road unless the whole frame is road.
2. If an annual crop is identifiable on the ground → that crop, even if trees stand in FRONT or BEHIND.
3. Planted orchard beats weedy understory.
4. Do not invent extra classes (no fallow, no tree savanna, no agroforestry, no "other" except Unclassifiable).
5. If the ground is hidden or two land uses are co-dominant → Unclassifiable, not Shrubby vegetation.
JSON only:
{"class_name_en":"<exact label>","confidence":<0-1>,"cues":["<cue>"],"reasoning":"<1-2 sentences>","usable_for_satellite_gt":<true|false>,"reject_reason":null|"unusable_image"|"ground_not_visible"|"mixed_plots"|"unknown_land_use"}