kangnicui2 commited on
Commit
710f363
·
verified ·
1 Parent(s): b54221f

Add dataset card with data details, split, class schema, and Croissant notes

Browse files
Files changed (1) hide show
  1. README.md +108 -0
README.md ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: GoldMiningMDD
3
+ license: other
4
+ task_categories:
5
+ - image-segmentation
6
+ task_ids:
7
+ - semantic-segmentation
8
+ language:
9
+ - en
10
+ tags:
11
+ - remote-sensing
12
+ - geospatial
13
+ - drone-imagery
14
+ - semantic-segmentation
15
+ size_categories:
16
+ - n<1K
17
+ ---
18
+
19
+ # GoldMiningMDD
20
+
21
+ GoldMiningMDD is a drone orthomosaic semantic-segmentation dataset for mining-site analysis in Madre de Dios, Peru.
22
+
23
+ - 12 sites
24
+ - 24 PNG files total (12 images + 12 labels)
25
+ - Splits: train/val/test at site level
26
+ - Labels: 14 foreground classes (`1..14`) + background (`0`)
27
+
28
+ ## Directory structure
29
+
30
+ ```text
31
+ train/
32
+ image/*.png
33
+ label/*.png
34
+ val/
35
+ image/*.png
36
+ label/*.png
37
+ test/
38
+ image/*.png
39
+ label/*.png
40
+ ```
41
+
42
+ ## Split definition
43
+
44
+ - **train (4 sites):** `AcumulacionAaron2B`, `Kotsimba`, `Los5Rebeldes`, `PlayaMirador1`
45
+ - **val (3 sites):** `Clavelito`, `Nayda`, `PlayaMirador2`
46
+ - **test (5 sites):** `Anel`, `ElEngano`, `Linda`, `Paolita`, `SantaInesDosMil`
47
+
48
+ `PlayaMirador` is split into two non-overlapping halves:
49
+ - `PlayaMirador1`: bottom half (train)
50
+ - `PlayaMirador2`: top half (val)
51
+
52
+ ## Label schema (merged classes)
53
+
54
+ Class IDs are globally consistent across all sites.
55
+
56
+ | ID | Class |
57
+ |---:|---|
58
+ | 0 | Background |
59
+ | 1 | Building |
60
+ | 2 | Mining raft |
61
+ | 3 | Primary Forest |
62
+ | 4 | Heavy machinery |
63
+ | 5 | Water bodies |
64
+ | 6 | Agricultural crop |
65
+ | 7 | Compact mounds |
66
+ | 8 | Gravel mounds |
67
+ | 9 | Grass |
68
+ | 10 | Type 1 natural regeneration |
69
+ | 11 | Type 2 natural regeneration |
70
+ | 12 | Bare ground |
71
+ | 13 | Sluice |
72
+ | 14 | Vehicles |
73
+
74
+ ### Merge rules from original 19-class labels
75
+
76
+ - `Building` = original IDs `{1 (Urban area), 4 (Mining camp)}`
77
+ - `Heavy machinery` = `{5 (Front loader), 8 (Excavator), 9 (Heavy machinery), 19 (Dump truck)}`
78
+ - `Vehicles` = `{17 (Vehicles), 18 (Small vehicles)}`
79
+ - Other classes map 1-to-1 to their corresponding semantic category.
80
+
81
+ ## Image-label pairing
82
+
83
+ - Pairing is by basename under each split:
84
+ - `train/image/AcumulacionAaron2B.png` ↔ `train/label/AcumulacionAaron2B.png`
85
+ - Label rasters are aligned to image spatial extent and grid before export.
86
+
87
+ ## Spatial/reference notes
88
+
89
+ - Source CRS in preprocessing: EPSG:32719
90
+ - This repository stores aligned PNG exports for training/evaluation.
91
+ - Full spatial metadata and preprocessing details are maintained in the project-side documentation.
92
+
93
+ ## Recommended usage
94
+
95
+ - Treat label value `0` as background/ignore if your training protocol does not score background.
96
+ - Keep site-level split boundaries (no cross-split mixing of sites).
97
+
98
+ ## Croissant status
99
+
100
+ The Hub `/croissant` endpoint currently returns unavailable for this dataset (no Hub-parquet representation yet).
101
+
102
+ Endpoint:
103
+
104
+ ```text
105
+ https://huggingface.co/api/datasets/kangnicui2/GoldMiningMDD/croissant
106
+ ```
107
+
108
+ To get auto-generated Croissant on Hugging Face, publish the dataset in a form that the Dataset Viewer can convert to Parquet (or already provides Parquet). Then the `/croissant` endpoint is generated automatically by the Hub.