Spaces:
Running on Zero
Running on Zero
Upgrade to professional LULC workbench
Browse files- .codex-plugin/plugin.json +8 -8
- .gitignore +0 -1
- LICENSE +0 -1
- README.md +32 -6
- app.py +331 -114
- satellite_utils.py +110 -1
- scripts/satellite_client.py +20 -4
- skills/analyze-satellite-imagery/SKILL.md +14 -8
- skills/analyze-satellite-imagery/agents/openai.yaml +2 -2
- skills/analyze-satellite-imagery/references/model-guide.md +13 -1
- tests/test_app_contract.py +5 -2
- tests/test_satellite_utils.py +17 -1
.codex-plugin/plugin.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
{
|
| 2 |
"name": "satellite-vision-toolkit",
|
| 3 |
-
"version": "0.
|
| 4 |
-
"description": "
|
| 5 |
"author": {
|
| 6 |
"name": "Mingze Chen",
|
| 7 |
"url": "https://github.com/LabMingzeChen"
|
|
@@ -12,16 +12,16 @@
|
|
| 12 |
"skills": "./skills/",
|
| 13 |
"interface": {
|
| 14 |
"displayName": "Satellite Vision Toolkit",
|
| 15 |
-
"shortDescription": "
|
| 16 |
-
"longDescription": "Analyze satellite and aerial images with
|
| 17 |
"developerName": "Mingze Chen",
|
| 18 |
"category": "Developer Tools",
|
| 19 |
-
"capabilities": ["Analyze", "Export"],
|
| 20 |
"defaultPrompt": [
|
| 21 |
-
"
|
| 22 |
-
"
|
|
|
|
| 23 |
"Explain the limits of these remote-sensing results."
|
| 24 |
]
|
| 25 |
}
|
| 26 |
}
|
| 27 |
-
|
|
|
|
| 1 |
{
|
| 2 |
"name": "satellite-vision-toolkit",
|
| 3 |
+
"version": "0.2.0",
|
| 4 |
+
"description": "Classify land use and land cover, segment surface classes, and detect objects in satellite and aerial imagery.",
|
| 5 |
"author": {
|
| 6 |
"name": "Mingze Chen",
|
| 7 |
"url": "https://github.com/LabMingzeChen"
|
|
|
|
| 12 |
"skills": "./skills/",
|
| 13 |
"interface": {
|
| 14 |
"displayName": "Satellite Vision Toolkit",
|
| 15 |
+
"shortDescription": "Professional multi-level analysis for overhead imagery.",
|
| 16 |
+
"longDescription": "Analyze satellite and aerial images with scene-level EuroSAT LULC classification, OpenEarthMap pixel segmentation, and remote-sensing object detection, then export visual, tabular, JSON, and pixel-coordinate evidence.",
|
| 17 |
"developerName": "Mingze Chen",
|
| 18 |
"category": "Developer Tools",
|
| 19 |
+
"capabilities": ["Classify", "Analyze", "Export"],
|
| 20 |
"defaultPrompt": [
|
| 21 |
+
"Run a complete professional assessment of this satellite image.",
|
| 22 |
+
"Classify the scene into land-use and land-cover categories.",
|
| 23 |
+
"Segment land cover and detect supported objects.",
|
| 24 |
"Explain the limits of these remote-sensing results."
|
| 25 |
]
|
| 26 |
}
|
| 27 |
}
|
|
|
.gitignore
CHANGED
|
@@ -4,4 +4,3 @@ __pycache__/
|
|
| 4 |
.venv/
|
| 5 |
.DS_Store
|
| 6 |
outputs/
|
| 7 |
-
|
|
|
|
| 4 |
.venv/
|
| 5 |
.DS_Store
|
| 6 |
outputs/
|
|
|
LICENSE
CHANGED
|
@@ -19,4 +19,3 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
| 19 |
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
SOFTWARE.
|
| 22 |
-
|
|
|
|
| 19 |
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
SOFTWARE.
|
|
|
README.md
CHANGED
|
@@ -9,21 +9,22 @@ python_version: "3.11"
|
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
license: mit
|
| 12 |
-
short_description:
|
| 13 |
---
|
| 14 |
|
| 15 |
<div align="center">
|
| 16 |
|
| 17 |
# 🛰️ Satellite Vision Toolkit
|
| 18 |
|
| 19 |
-
###
|
| 20 |
|
| 21 |
[](https://huggingface.co/spaces/Mingze/SatelliteVisionToolkit)
|
| 22 |
[](https://huggingface.co/bluelabel/satellite-equipment-detection-yolov8n-vhr10)
|
| 23 |
[](https://huggingface.co/mfaytin/mask2former-satellite)
|
|
|
|
| 24 |
[](LICENSE)
|
| 25 |
|
| 26 |
-
**Upload one satellite or aerial image,
|
| 27 |
|
| 28 |
[**🚀 Launch the live app**](https://huggingface.co/spaces/Mingze/SatelliteVisionToolkit) ·
|
| 29 |
[**💻 GitHub source**](https://github.com/LabMingzeChen/SatelliteVisionToolkit)
|
|
@@ -32,24 +33,32 @@ short_description: Satellite object detection and land-cover segmentation.
|
|
| 32 |
|
| 33 |
## What it does
|
| 34 |
|
| 35 |
-
The app provides
|
| 36 |
|
| 37 |
| Mode | Model | Output vocabulary |
|
| 38 |
|---|---|---|
|
|
|
|
| 39 |
| Object detection | YOLOv8n fine-tuned on NWPU VHR-10 | airplane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, vehicle |
|
| 40 |
| Land-cover segmentation | Mask2Former fine-tuned on OpenEarthMap | background, bare land, grass, pavement, road, tree, water, cropland, building |
|
| 41 |
|
| 42 |
Every run creates visual and machine-readable outputs:
|
| 43 |
|
|
|
|
| 44 |
- Detection overlay, per-class summary, per-object CSV, and pixel-coordinate GeoJSON.
|
| 45 |
- Land-cover overlay, categorical mask, raw class-ID PNG, and class-area CSV.
|
| 46 |
-
-
|
|
|
|
| 47 |
- A reusable Codex skill and command-line Space client.
|
| 48 |
|
| 49 |
## How it works
|
| 50 |
|
| 51 |
```text
|
| 52 |
Satellite or aerial RGB image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
├── YOLOv8n / NWPU VHR-10
|
| 54 |
│ ├── labeled bounding-box overlay
|
| 55 |
│ ├── class counts and confidence
|
|
@@ -85,6 +94,12 @@ from gradio_client import Client, handle_file
|
|
| 85 |
|
| 86 |
client = Client("Mingze/SatelliteVisionToolkit")
|
| 87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
detection = client.predict(
|
| 89 |
handle_file("satellite.jpg"),
|
| 90 |
0.25,
|
|
@@ -98,13 +113,21 @@ segmentation = client.predict(
|
|
| 98 |
0.10,
|
| 99 |
api_name="/segment",
|
| 100 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
```
|
| 102 |
|
| 103 |
The bundled CLI wraps the same endpoints:
|
| 104 |
|
| 105 |
```bash
|
|
|
|
| 106 |
python scripts/satellite_client.py detect satellite.jpg --output detection.json
|
| 107 |
python scripts/satellite_client.py segment satellite.jpg --output segmentation.json
|
|
|
|
| 108 |
```
|
| 109 |
|
| 110 |
## Project structure
|
|
@@ -125,6 +148,8 @@ The application code is MIT licensed. Model software, weights, and training data
|
|
| 125 |
|
| 126 |
| Resource | Role | Terms noted by source |
|
| 127 |
|---|---|---|
|
|
|
|
|
|
|
| 128 |
| [`bluelabel/satellite-equipment-detection-yolov8n-vhr10`](https://huggingface.co/bluelabel/satellite-equipment-detection-yolov8n-vhr10) | Remote-sensing object detector | Model card lists MIT; Ultralytics runtime has separate licensing |
|
| 129 |
| [NWPU VHR-10](https://gcheng-nwpu.github.io/#Datasets) | Detection training dataset | Review dataset terms and cite its authors |
|
| 130 |
| [`mfaytin/mask2former-satellite`](https://huggingface.co/mfaytin/mask2former-satellite) | Land-cover segmentation model | Model card lists MIT |
|
|
@@ -133,6 +158,8 @@ The application code is MIT licensed. Model software, weights, and training data
|
|
| 133 |
## Limitations and responsible use
|
| 134 |
|
| 135 |
- Results vary with spatial resolution, sensor, geography, season, atmospheric conditions, shadows, and image preprocessing.
|
|
|
|
|
|
|
| 136 |
- Small objects may disappear during resizing or fall below the confidence threshold.
|
| 137 |
- Detection counts describe visible predictions, not complete inventories.
|
| 138 |
- Segmentation shares describe processed image pixels, not surveyed ground area.
|
|
@@ -152,4 +179,3 @@ The application code is MIT licensed. Model software, weights, and training data
|
|
| 152 |
```
|
| 153 |
|
| 154 |
Please also cite NWPU VHR-10, OpenEarthMap, YOLO/Ultralytics, and Mask2Former as applicable.
|
| 155 |
-
|
|
|
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
license: mit
|
| 12 |
+
short_description: LULC classification, segmentation, and object detection.
|
| 13 |
---
|
| 14 |
|
| 15 |
<div align="center">
|
| 16 |
|
| 17 |
# 🛰️ Satellite Vision Toolkit
|
| 18 |
|
| 19 |
+
### Professional scene, pixel, and object-level analysis for overhead imagery
|
| 20 |
|
| 21 |
[](https://huggingface.co/spaces/Mingze/SatelliteVisionToolkit)
|
| 22 |
[](https://huggingface.co/bluelabel/satellite-equipment-detection-yolov8n-vhr10)
|
| 23 |
[](https://huggingface.co/mfaytin/mask2former-satellite)
|
| 24 |
+
[](https://huggingface.co/mrm8488/convnext-tiny-finetuned-eurosat)
|
| 25 |
[](LICENSE)
|
| 26 |
|
| 27 |
+
**Upload one satellite or aerial image, run a multi-model assessment, and download reusable visual and machine-readable evidence.**
|
| 28 |
|
| 29 |
[**🚀 Launch the live app**](https://huggingface.co/spaces/Mingze/SatelliteVisionToolkit) ·
|
| 30 |
[**💻 GitHub source**](https://github.com/LabMingzeChen/SatelliteVisionToolkit)
|
|
|
|
| 33 |
|
| 34 |
## What it does
|
| 35 |
|
| 36 |
+
The app provides three complementary analytical levels plus a one-click combined workflow:
|
| 37 |
|
| 38 |
| Mode | Model | Output vocabulary |
|
| 39 |
|---|---|---|
|
| 40 |
+
| Scene-level LULC classification | ConvNeXT-Tiny fine-tuned on EuroSAT | annual crop, forest, herbaceous vegetation, highway, industrial, pasture, permanent crop, residential, river, sea/lake |
|
| 41 |
| Object detection | YOLOv8n fine-tuned on NWPU VHR-10 | airplane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, vehicle |
|
| 42 |
| Land-cover segmentation | Mask2Former fine-tuned on OpenEarthMap | background, bare land, grass, pavement, road, tree, water, cropland, building |
|
| 43 |
|
| 44 |
Every run creates visual and machine-readable outputs:
|
| 45 |
|
| 46 |
+
- Ranked LULC probabilities, a confidence tier, normalized entropy, and CSV/JSON exports.
|
| 47 |
- Detection overlay, per-class summary, per-object CSV, and pixel-coordinate GeoJSON.
|
| 48 |
- Land-cover overlay, categorical mask, raw class-ID PNG, and class-area CSV.
|
| 49 |
+
- A professional executive dashboard and complete JSON evidence package.
|
| 50 |
+
- Independent Gradio API endpoints at `/classify`, `/segment`, `/detect`, and `/analyze`.
|
| 51 |
- A reusable Codex skill and command-line Space client.
|
| 52 |
|
| 53 |
## How it works
|
| 54 |
|
| 55 |
```text
|
| 56 |
Satellite or aerial RGB image
|
| 57 |
+
├── ConvNeXT-Tiny / EuroSAT
|
| 58 |
+
│ ├── ranked scene-level LULC probabilities
|
| 59 |
+
│ ├── normalized uncertainty (entropy)
|
| 60 |
+
│ └── classification CSV + JSON
|
| 61 |
+
│
|
| 62 |
├── YOLOv8n / NWPU VHR-10
|
| 63 |
│ ├── labeled bounding-box overlay
|
| 64 |
│ ├── class counts and confidence
|
|
|
|
| 94 |
|
| 95 |
client = Client("Mingze/SatelliteVisionToolkit")
|
| 96 |
|
| 97 |
+
classification = client.predict(
|
| 98 |
+
handle_file("satellite.jpg"),
|
| 99 |
+
5,
|
| 100 |
+
api_name="/classify",
|
| 101 |
+
)
|
| 102 |
+
|
| 103 |
detection = client.predict(
|
| 104 |
handle_file("satellite.jpg"),
|
| 105 |
0.25,
|
|
|
|
| 113 |
0.10,
|
| 114 |
api_name="/segment",
|
| 115 |
)
|
| 116 |
+
|
| 117 |
+
complete = client.predict(
|
| 118 |
+
handle_file("satellite.jpg"),
|
| 119 |
+
5, 0.55, 0.10, 0.25, 0.45,
|
| 120 |
+
api_name="/analyze",
|
| 121 |
+
)
|
| 122 |
```
|
| 123 |
|
| 124 |
The bundled CLI wraps the same endpoints:
|
| 125 |
|
| 126 |
```bash
|
| 127 |
+
python scripts/satellite_client.py classify satellite.jpg --output classification.json
|
| 128 |
python scripts/satellite_client.py detect satellite.jpg --output detection.json
|
| 129 |
python scripts/satellite_client.py segment satellite.jpg --output segmentation.json
|
| 130 |
+
python scripts/satellite_client.py analyze satellite.jpg --output complete.json
|
| 131 |
```
|
| 132 |
|
| 133 |
## Project structure
|
|
|
|
| 148 |
|
| 149 |
| Resource | Role | Terms noted by source |
|
| 150 |
|---|---|---|
|
| 151 |
+
| [`mrm8488/convnext-tiny-finetuned-eurosat`](https://huggingface.co/mrm8488/convnext-tiny-finetuned-eurosat) | Scene-level LULC classifier | Model card lists Apache-2.0 |
|
| 152 |
+
| [EuroSAT](https://huggingface.co/datasets/GFM-Bench/EuroSAT) | LULC classification dataset | Review dataset terms and cite Helber et al. |
|
| 153 |
| [`bluelabel/satellite-equipment-detection-yolov8n-vhr10`](https://huggingface.co/bluelabel/satellite-equipment-detection-yolov8n-vhr10) | Remote-sensing object detector | Model card lists MIT; Ultralytics runtime has separate licensing |
|
| 154 |
| [NWPU VHR-10](https://gcheng-nwpu.github.io/#Datasets) | Detection training dataset | Review dataset terms and cite its authors |
|
| 155 |
| [`mfaytin/mask2former-satellite`](https://huggingface.co/mfaytin/mask2former-satellite) | Land-cover segmentation model | Model card lists MIT |
|
|
|
|
| 158 |
## Limitations and responsible use
|
| 159 |
|
| 160 |
- Results vary with spatial resolution, sensor, geography, season, atmospheric conditions, shadows, and image preprocessing.
|
| 161 |
+
- EuroSAT classification is a whole-scene hypothesis learned from small European Sentinel-2 RGB tiles; it is not parcel delineation, zoning, cadastral, or legal land-use evidence.
|
| 162 |
+
- Review ranked alternatives and normalized entropy. A confident prediction can still be wrong under domain shift.
|
| 163 |
- Small objects may disappear during resizing or fall below the confidence threshold.
|
| 164 |
- Detection counts describe visible predictions, not complete inventories.
|
| 165 |
- Segmentation shares describe processed image pixels, not surveyed ground area.
|
|
|
|
| 179 |
```
|
| 180 |
|
| 181 |
Please also cite NWPU VHR-10, OpenEarthMap, YOLO/Ultralytics, and Mask2Former as applicable.
|
|
|
app.py
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
"""Hugging Face Space for satellite
|
| 2 |
|
| 3 |
from __future__ import annotations
|
| 4 |
|
|
@@ -13,7 +13,11 @@ import numpy as np
|
|
| 13 |
import torch
|
| 14 |
from huggingface_hub import hf_hub_download
|
| 15 |
from PIL import Image
|
| 16 |
-
from transformers import
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
try:
|
| 19 |
import spaces
|
|
@@ -27,18 +31,25 @@ except ImportError:
|
|
| 27 |
spaces = _SpacesFallback()
|
| 28 |
|
| 29 |
from satellite_utils import (
|
|
|
|
| 30 |
build_class_table,
|
| 31 |
build_detection_summary,
|
| 32 |
build_detection_table,
|
|
|
|
|
|
|
| 33 |
render_detections,
|
|
|
|
| 34 |
render_segmentation,
|
| 35 |
resize_for_inference,
|
| 36 |
write_class_csv,
|
| 37 |
write_detection_csv,
|
|
|
|
|
|
|
| 38 |
write_pixel_geojson,
|
| 39 |
)
|
| 40 |
|
| 41 |
|
|
|
|
| 42 |
SEGMENTATION_MODEL_ID = "mfaytin/mask2former-satellite"
|
| 43 |
DETECTION_MODEL_ID = "bluelabel/satellite-equipment-detection-yolov8n-vhr10"
|
| 44 |
DETECTION_FILENAME = "best.pt"
|
|
@@ -56,16 +67,25 @@ OPEN_EARTH_MAP_LABELS = {
|
|
| 56 |
}
|
| 57 |
|
| 58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
@lru_cache(maxsize=1)
|
| 60 |
def load_segmenter():
|
| 61 |
-
|
| 62 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 63 |
processor = AutoImageProcessor.from_pretrained(SEGMENTATION_MODEL_ID)
|
| 64 |
-
model = (
|
| 65 |
-
Mask2FormerForUniversalSegmentation.from_pretrained(SEGMENTATION_MODEL_ID)
|
| 66 |
-
.to(device)
|
| 67 |
-
.eval()
|
| 68 |
-
)
|
| 69 |
return processor, model, OPEN_EARTH_MAP_LABELS, device
|
| 70 |
|
| 71 |
|
|
@@ -83,34 +103,62 @@ def _new_output_dir() -> Path:
|
|
| 83 |
return output_dir
|
| 84 |
|
| 85 |
|
| 86 |
-
|
| 87 |
-
def segment_satellite_image(
|
| 88 |
-
image: Image.Image | None,
|
| 89 |
-
opacity: float,
|
| 90 |
-
min_share_percent: float,
|
| 91 |
-
):
|
| 92 |
if image is None:
|
| 93 |
raise gr.Error("Please upload a satellite or aerial image first.")
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
try:
|
| 97 |
-
processor, model, id2label, device = load_segmenter()
|
| 98 |
-
inputs = processor(images=prepared, return_tensors="pt")
|
| 99 |
-
inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
|
| 100 |
-
with torch.inference_mode():
|
| 101 |
-
outputs = model(**inputs)
|
| 102 |
-
class_map = processor.post_process_semantic_segmentation(
|
| 103 |
-
outputs,
|
| 104 |
-
target_sizes=[(prepared.height, prepared.width)],
|
| 105 |
-
)[0].cpu().numpy().astype(np.uint8)
|
| 106 |
-
except Exception as exc:
|
| 107 |
-
raise gr.Error(f"Land-cover segmentation failed: {type(exc).__name__}: {exc}") from exc
|
| 108 |
|
| 109 |
-
|
| 110 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
rows = build_class_table(class_map, id2label, float(min_share_percent))
|
| 113 |
-
output_dir = _new_output_dir()
|
| 114 |
overlay_path = output_dir / "land_cover_overlay.png"
|
| 115 |
mask_path = output_dir / "land_cover_color_mask.png"
|
| 116 |
ids_path = output_dir / "land_cover_class_ids.png"
|
|
@@ -119,110 +167,250 @@ def segment_satellite_image(
|
|
| 119 |
color_mask.save(mask_path)
|
| 120 |
Image.fromarray(class_map).save(ids_path)
|
| 121 |
write_class_csv(csv_path, rows)
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
|
|
|
| 128 |
|
| 129 |
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
image: Image.Image | None,
|
| 133 |
confidence_threshold: float,
|
| 134 |
iou_threshold: float,
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
raise gr.Error("Please upload a satellite or aerial image first.")
|
| 138 |
-
started_at = time.perf_counter()
|
| 139 |
-
prepared = resize_for_inference(image)
|
| 140 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
)
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
)
|
| 171 |
-
except Exception as exc:
|
| 172 |
-
raise gr.Error(f"Satellite object detection failed: {type(exc).__name__}: {exc}") from exc
|
| 173 |
-
|
| 174 |
overlay = render_detections(prepared, detections)
|
| 175 |
summary_rows = build_detection_summary(detections)
|
| 176 |
detail_rows = build_detection_table(detections, prepared.size)
|
| 177 |
-
output_dir = _new_output_dir()
|
| 178 |
overlay_path = output_dir / "satellite_detection_overlay.png"
|
| 179 |
csv_path = output_dir / "satellite_detections.csv"
|
| 180 |
geojson_path = output_dir / "satellite_detections_pixel_coordinates.geojson"
|
| 181 |
overlay.save(overlay_path)
|
| 182 |
write_detection_csv(csv_path, detail_rows)
|
| 183 |
write_pixel_geojson(geojson_path, detections, prepared.size)
|
| 184 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
status = (
|
| 186 |
-
f"
|
| 187 |
-
f"{
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
)
|
| 189 |
-
return overlay, summary_rows, detail_rows, [str(overlay_path), str(csv_path), str(geojson_path)], status
|
| 190 |
|
| 191 |
|
| 192 |
CSS = """
|
| 193 |
-
.gradio-container {max-width:
|
| 194 |
-
.hero {
|
| 195 |
-
.hero h1 {font-size: 2.
|
| 196 |
-
.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 197 |
"""
|
| 198 |
|
| 199 |
-
|
|
|
|
| 200 |
gr.HTML("""
|
| 201 |
<div class="hero">
|
| 202 |
-
<
|
| 203 |
-
<
|
| 204 |
-
<p>
|
| 205 |
-
<a href="https://huggingface.co/mfaytin/mask2former-satellite">Segmentation model</a> ·
|
| 206 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
</div>
|
| 208 |
""")
|
| 209 |
-
with gr.Row():
|
| 210 |
-
image_input = gr.Image(type="pil", label="Satellite / aerial image", height=
|
| 211 |
with gr.Column():
|
| 212 |
-
gr.Markdown(""
|
| 213 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
|
| 215 |
-
|
| 216 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
|
| 218 |
-
|
| 219 |
-
""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 220 |
|
| 221 |
-
with gr.Tabs():
|
| 222 |
with gr.Tab("Land-cover segmentation"):
|
| 223 |
-
|
| 224 |
-
opacity = gr.Slider(0.1, 0.9, value=0.55, step=0.05, label="Overlay opacity")
|
| 225 |
-
min_share = gr.Slider(0.0, 5.0, value=0.1, step=0.1, label="Minimum table share (%)")
|
| 226 |
segment_button = gr.Button("Segment land cover", variant="primary")
|
| 227 |
segment_status = gr.Markdown()
|
| 228 |
with gr.Row():
|
|
@@ -230,22 +418,18 @@ RGB PNG/JPEG/WebP/TIFF images work best. Results are image-space estimates, not
|
|
| 230 |
segment_mask = gr.Image(label="Categorical mask")
|
| 231 |
segment_table = gr.Dataframe(
|
| 232 |
headers=["Class ID", "Class", "Pixels", "Share (%)", "Color"],
|
| 233 |
-
datatype=["number", "str", "number", "number", "str"],
|
| 234 |
interactive=False,
|
| 235 |
label="Land-cover area summary",
|
| 236 |
)
|
| 237 |
segment_files = gr.File(label="Download segmentation outputs", file_count="multiple")
|
| 238 |
|
| 239 |
with gr.Tab("Object detection"):
|
| 240 |
-
|
| 241 |
-
confidence = gr.Slider(0.05, 0.9, value=0.25, step=0.05, label="Confidence threshold")
|
| 242 |
-
iou = gr.Slider(0.1, 0.9, value=0.45, step=0.05, label="NMS IoU threshold")
|
| 243 |
detect_button = gr.Button("Detect satellite objects", variant="primary")
|
| 244 |
detect_status = gr.Markdown()
|
| 245 |
detect_overlay = gr.Image(label="Detection overlay")
|
| 246 |
detection_summary = gr.Dataframe(
|
| 247 |
headers=["Class", "Count", "Average confidence", "Maximum confidence"],
|
| 248 |
-
datatype=["str", "number", "number", "number"],
|
| 249 |
interactive=False,
|
| 250 |
label="Detection summary",
|
| 251 |
)
|
|
@@ -256,10 +440,25 @@ RGB PNG/JPEG/WebP/TIFF images work best. Results are image-space estimates, not
|
|
| 256 |
)
|
| 257 |
detection_files = gr.File(label="Download detection outputs", file_count="multiple")
|
| 258 |
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 262 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 263 |
segment_button.click(
|
| 264 |
segment_satellite_image,
|
| 265 |
inputs=[image_input, opacity, min_share],
|
|
@@ -272,6 +471,24 @@ RGB PNG/JPEG/WebP/TIFF images work best. Results are image-space estimates, not
|
|
| 272 |
outputs=[detect_overlay, detection_summary, detection_details, detection_files, detect_status],
|
| 273 |
api_name="detect",
|
| 274 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 275 |
|
| 276 |
|
| 277 |
if __name__ == "__main__":
|
|
|
|
| 1 |
+
"""Professional Hugging Face Space for multi-level satellite image analysis."""
|
| 2 |
|
| 3 |
from __future__ import annotations
|
| 4 |
|
|
|
|
| 13 |
import torch
|
| 14 |
from huggingface_hub import hf_hub_download
|
| 15 |
from PIL import Image
|
| 16 |
+
from transformers import (
|
| 17 |
+
AutoImageProcessor,
|
| 18 |
+
AutoModelForImageClassification,
|
| 19 |
+
Mask2FormerForUniversalSegmentation,
|
| 20 |
+
)
|
| 21 |
|
| 22 |
try:
|
| 23 |
import spaces
|
|
|
|
| 31 |
spaces = _SpacesFallback()
|
| 32 |
|
| 33 |
from satellite_utils import (
|
| 34 |
+
build_analysis_summary,
|
| 35 |
build_class_table,
|
| 36 |
build_detection_summary,
|
| 37 |
build_detection_table,
|
| 38 |
+
build_lulc_table,
|
| 39 |
+
normalized_entropy,
|
| 40 |
render_detections,
|
| 41 |
+
render_lulc_assessment,
|
| 42 |
render_segmentation,
|
| 43 |
resize_for_inference,
|
| 44 |
write_class_csv,
|
| 45 |
write_detection_csv,
|
| 46 |
+
write_json,
|
| 47 |
+
write_lulc_csv,
|
| 48 |
write_pixel_geojson,
|
| 49 |
)
|
| 50 |
|
| 51 |
|
| 52 |
+
CLASSIFICATION_MODEL_ID = "mrm8488/convnext-tiny-finetuned-eurosat"
|
| 53 |
SEGMENTATION_MODEL_ID = "mfaytin/mask2former-satellite"
|
| 54 |
DETECTION_MODEL_ID = "bluelabel/satellite-equipment-detection-yolov8n-vhr10"
|
| 55 |
DETECTION_FILENAME = "best.pt"
|
|
|
|
| 67 |
}
|
| 68 |
|
| 69 |
|
| 70 |
+
def _device() -> torch.device:
|
| 71 |
+
torch.set_num_threads(max(1, min(4, os.cpu_count() or 1)))
|
| 72 |
+
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
@lru_cache(maxsize=1)
|
| 76 |
+
def load_classifier():
|
| 77 |
+
device = _device()
|
| 78 |
+
processor = AutoImageProcessor.from_pretrained(CLASSIFICATION_MODEL_ID)
|
| 79 |
+
model = AutoModelForImageClassification.from_pretrained(CLASSIFICATION_MODEL_ID).to(device).eval()
|
| 80 |
+
id2label = {int(key): str(value) for key, value in model.config.id2label.items()}
|
| 81 |
+
return processor, model, id2label, device
|
| 82 |
+
|
| 83 |
+
|
| 84 |
@lru_cache(maxsize=1)
|
| 85 |
def load_segmenter():
|
| 86 |
+
device = _device()
|
|
|
|
| 87 |
processor = AutoImageProcessor.from_pretrained(SEGMENTATION_MODEL_ID)
|
| 88 |
+
model = Mask2FormerForUniversalSegmentation.from_pretrained(SEGMENTATION_MODEL_ID).to(device).eval()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
return processor, model, OPEN_EARTH_MAP_LABELS, device
|
| 90 |
|
| 91 |
|
|
|
|
| 103 |
return output_dir
|
| 104 |
|
| 105 |
|
| 106 |
+
def _require_image(image: Image.Image | None) -> Image.Image:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
if image is None:
|
| 108 |
raise gr.Error("Please upload a satellite or aerial image first.")
|
| 109 |
+
return resize_for_inference(image)
|
| 110 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
+
def _classify_impl(prepared: Image.Image, top_k: int, output_dir: Path) -> dict[str, object]:
|
| 113 |
+
processor, model, id2label, device = load_classifier()
|
| 114 |
+
inputs = {name: tensor.to(device) for name, tensor in processor(images=prepared, return_tensors="pt").items()}
|
| 115 |
+
with torch.inference_mode():
|
| 116 |
+
logits = model(**inputs).logits[0]
|
| 117 |
+
probabilities = torch.softmax(logits, dim=-1).detach().cpu().tolist()
|
| 118 |
+
rows = build_lulc_table(probabilities, id2label, top_k)
|
| 119 |
+
entropy = normalized_entropy(probabilities)
|
| 120 |
+
csv_path = output_dir / "lulc_classification.csv"
|
| 121 |
+
json_path = output_dir / "lulc_classification.json"
|
| 122 |
+
write_lulc_csv(csv_path, rows)
|
| 123 |
+
write_json(
|
| 124 |
+
json_path,
|
| 125 |
+
{
|
| 126 |
+
"task": "scene_level_lulc_classification",
|
| 127 |
+
"model": CLASSIFICATION_MODEL_ID,
|
| 128 |
+
"processed_image_size": {"width": prepared.width, "height": prepared.height},
|
| 129 |
+
"normalized_entropy": round(entropy, 6),
|
| 130 |
+
"predictions": [
|
| 131 |
+
{"rank": row[0], "class": row[1], "probability_percent": row[2], "confidence_tier": row[3]}
|
| 132 |
+
for row in rows
|
| 133 |
+
],
|
| 134 |
+
"scope_note": "Whole-scene EuroSAT class; not a cadastral or planning land-use designation.",
|
| 135 |
+
},
|
| 136 |
)
|
| 137 |
+
return {
|
| 138 |
+
"rows": rows,
|
| 139 |
+
"entropy": entropy,
|
| 140 |
+
"assessment": render_lulc_assessment(rows, entropy),
|
| 141 |
+
"files": [str(csv_path), str(json_path)],
|
| 142 |
+
"device": device.type,
|
| 143 |
+
}
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
def _segment_impl(
|
| 147 |
+
prepared: Image.Image,
|
| 148 |
+
opacity: float,
|
| 149 |
+
min_share_percent: float,
|
| 150 |
+
output_dir: Path,
|
| 151 |
+
) -> dict[str, object]:
|
| 152 |
+
processor, model, id2label, device = load_segmenter()
|
| 153 |
+
inputs = {name: tensor.to(device) for name, tensor in processor(images=prepared, return_tensors="pt").items()}
|
| 154 |
+
with torch.inference_mode():
|
| 155 |
+
outputs = model(**inputs)
|
| 156 |
+
class_map = processor.post_process_semantic_segmentation(
|
| 157 |
+
outputs,
|
| 158 |
+
target_sizes=[(prepared.height, prepared.width)],
|
| 159 |
+
)[0].cpu().numpy().astype(np.uint8)
|
| 160 |
+
overlay, color_mask = render_segmentation(prepared, class_map, id2label, float(opacity))
|
| 161 |
rows = build_class_table(class_map, id2label, float(min_share_percent))
|
|
|
|
| 162 |
overlay_path = output_dir / "land_cover_overlay.png"
|
| 163 |
mask_path = output_dir / "land_cover_color_mask.png"
|
| 164 |
ids_path = output_dir / "land_cover_class_ids.png"
|
|
|
|
| 167 |
color_mask.save(mask_path)
|
| 168 |
Image.fromarray(class_map).save(ids_path)
|
| 169 |
write_class_csv(csv_path, rows)
|
| 170 |
+
return {
|
| 171 |
+
"overlay": overlay,
|
| 172 |
+
"mask": color_mask,
|
| 173 |
+
"rows": rows,
|
| 174 |
+
"files": [str(overlay_path), str(mask_path), str(ids_path), str(csv_path)],
|
| 175 |
+
"device": device.type,
|
| 176 |
+
}
|
| 177 |
|
| 178 |
|
| 179 |
+
def _detect_impl(
|
| 180 |
+
prepared: Image.Image,
|
|
|
|
| 181 |
confidence_threshold: float,
|
| 182 |
iou_threshold: float,
|
| 183 |
+
output_dir: Path,
|
| 184 |
+
) -> dict[str, object]:
|
|
|
|
|
|
|
|
|
|
| 185 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 186 |
+
detector = load_detector()
|
| 187 |
+
prediction = detector.predict(
|
| 188 |
+
source=np.asarray(prepared),
|
| 189 |
+
conf=float(confidence_threshold),
|
| 190 |
+
iou=float(iou_threshold),
|
| 191 |
+
imgsz=1024,
|
| 192 |
+
device=device,
|
| 193 |
+
max_det=500,
|
| 194 |
+
verbose=False,
|
| 195 |
+
)[0]
|
| 196 |
+
detections: list[dict[str, object]] = []
|
| 197 |
+
if prediction.boxes is not None:
|
| 198 |
+
for coordinates, confidence, class_id_value in zip(
|
| 199 |
+
prediction.boxes.xyxy.detach().cpu().tolist(),
|
| 200 |
+
prediction.boxes.conf.detach().cpu().tolist(),
|
| 201 |
+
prediction.boxes.cls.detach().cpu().tolist(),
|
| 202 |
+
):
|
| 203 |
+
class_id = int(class_id_value)
|
| 204 |
+
detections.append(
|
| 205 |
+
{
|
| 206 |
+
"class_id": class_id,
|
| 207 |
+
"class_name": str(prediction.names[class_id]),
|
| 208 |
+
"confidence": float(confidence),
|
| 209 |
+
"x1": float(coordinates[0]),
|
| 210 |
+
"y1": float(coordinates[1]),
|
| 211 |
+
"x2": float(coordinates[2]),
|
| 212 |
+
"y2": float(coordinates[3]),
|
| 213 |
+
}
|
| 214 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
overlay = render_detections(prepared, detections)
|
| 216 |
summary_rows = build_detection_summary(detections)
|
| 217 |
detail_rows = build_detection_table(detections, prepared.size)
|
|
|
|
| 218 |
overlay_path = output_dir / "satellite_detection_overlay.png"
|
| 219 |
csv_path = output_dir / "satellite_detections.csv"
|
| 220 |
geojson_path = output_dir / "satellite_detections_pixel_coordinates.geojson"
|
| 221 |
overlay.save(overlay_path)
|
| 222 |
write_detection_csv(csv_path, detail_rows)
|
| 223 |
write_pixel_geojson(geojson_path, detections, prepared.size)
|
| 224 |
+
return {
|
| 225 |
+
"overlay": overlay,
|
| 226 |
+
"summary": summary_rows,
|
| 227 |
+
"details": detail_rows,
|
| 228 |
+
"files": [str(overlay_path), str(csv_path), str(geojson_path)],
|
| 229 |
+
"device": device,
|
| 230 |
+
}
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
@spaces.GPU(duration=120)
|
| 234 |
+
def classify_lulc(image: Image.Image | None, top_k: int):
|
| 235 |
+
started_at = time.perf_counter()
|
| 236 |
+
prepared = _require_image(image)
|
| 237 |
+
try:
|
| 238 |
+
result = _classify_impl(prepared, int(top_k), _new_output_dir())
|
| 239 |
+
except Exception as exc:
|
| 240 |
+
raise gr.Error(f"LULC classification failed: {type(exc).__name__}: {exc}") from exc
|
| 241 |
+
status = (
|
| 242 |
+
f"Complete · {prepared.width}×{prepared.height} · top class {result['rows'][0][1]} "
|
| 243 |
+
f"({result['rows'][0][2]:.1f}%) · {time.perf_counter() - started_at:.1f}s · device={result['device']}"
|
| 244 |
+
)
|
| 245 |
+
return result["assessment"], result["rows"], result["files"], status
|
| 246 |
+
|
| 247 |
+
|
| 248 |
+
@spaces.GPU(duration=120)
|
| 249 |
+
def segment_satellite_image(image: Image.Image | None, opacity: float, min_share_percent: float):
|
| 250 |
+
started_at = time.perf_counter()
|
| 251 |
+
prepared = _require_image(image)
|
| 252 |
+
try:
|
| 253 |
+
result = _segment_impl(prepared, opacity, min_share_percent, _new_output_dir())
|
| 254 |
+
except Exception as exc:
|
| 255 |
+
raise gr.Error(f"Land-cover segmentation failed: {type(exc).__name__}: {exc}") from exc
|
| 256 |
status = (
|
| 257 |
+
f"Complete · {prepared.width}×{prepared.height} · {len(result['rows'])} reported cover classes · "
|
| 258 |
+
f"{time.perf_counter() - started_at:.1f}s · device={result['device']}"
|
| 259 |
+
)
|
| 260 |
+
return result["overlay"], result["mask"], result["rows"], result["files"], status
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
@spaces.GPU(duration=120)
|
| 264 |
+
def detect_satellite_objects(image: Image.Image | None, confidence_threshold: float, iou_threshold: float):
|
| 265 |
+
started_at = time.perf_counter()
|
| 266 |
+
prepared = _require_image(image)
|
| 267 |
+
try:
|
| 268 |
+
result = _detect_impl(prepared, confidence_threshold, iou_threshold, _new_output_dir())
|
| 269 |
+
except Exception as exc:
|
| 270 |
+
raise gr.Error(f"Satellite object detection failed: {type(exc).__name__}: {exc}") from exc
|
| 271 |
+
status = (
|
| 272 |
+
f"Complete · {prepared.width}×{prepared.height} · {len(result['details'])} objects · "
|
| 273 |
+
f"{len(result['summary'])} classes · {time.perf_counter() - started_at:.1f}s · device={result['device']}"
|
| 274 |
+
)
|
| 275 |
+
return result["overlay"], result["summary"], result["details"], result["files"], status
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
@spaces.GPU(duration=180)
|
| 279 |
+
def analyze_satellite_image(
|
| 280 |
+
image: Image.Image | None,
|
| 281 |
+
top_k: int,
|
| 282 |
+
opacity: float,
|
| 283 |
+
min_share_percent: float,
|
| 284 |
+
confidence_threshold: float,
|
| 285 |
+
iou_threshold: float,
|
| 286 |
+
):
|
| 287 |
+
started_at = time.perf_counter()
|
| 288 |
+
prepared = _require_image(image)
|
| 289 |
+
output_dir = _new_output_dir()
|
| 290 |
+
try:
|
| 291 |
+
classification = _classify_impl(prepared, int(top_k), output_dir)
|
| 292 |
+
segmentation = _segment_impl(prepared, opacity, min_share_percent, output_dir)
|
| 293 |
+
detection = _detect_impl(prepared, confidence_threshold, iou_threshold, output_dir)
|
| 294 |
+
except Exception as exc:
|
| 295 |
+
raise gr.Error(f"Complete analysis failed: {type(exc).__name__}: {exc}") from exc
|
| 296 |
+
elapsed = time.perf_counter() - started_at
|
| 297 |
+
summary = build_analysis_summary(
|
| 298 |
+
classification["rows"],
|
| 299 |
+
float(classification["entropy"]),
|
| 300 |
+
segmentation["rows"],
|
| 301 |
+
detection["summary"],
|
| 302 |
+
elapsed,
|
| 303 |
+
)
|
| 304 |
+
report_path = output_dir / "analysis_report.json"
|
| 305 |
+
write_json(
|
| 306 |
+
report_path,
|
| 307 |
+
{
|
| 308 |
+
"processed_image_size": {"width": prepared.width, "height": prepared.height},
|
| 309 |
+
"models": {
|
| 310 |
+
"classification": CLASSIFICATION_MODEL_ID,
|
| 311 |
+
"segmentation": SEGMENTATION_MODEL_ID,
|
| 312 |
+
"detection": DETECTION_MODEL_ID,
|
| 313 |
+
},
|
| 314 |
+
"lulc_classification": classification["rows"],
|
| 315 |
+
"lulc_normalized_entropy": round(float(classification["entropy"]), 6),
|
| 316 |
+
"land_cover_pixel_shares": segmentation["rows"],
|
| 317 |
+
"detection_summary": detection["summary"],
|
| 318 |
+
"detection_details": detection["details"],
|
| 319 |
+
"elapsed_seconds": round(elapsed, 3),
|
| 320 |
+
"coordinate_note": "Detection GeoJSON is in top-left-origin image pixels and has no geographic CRS.",
|
| 321 |
+
},
|
| 322 |
+
)
|
| 323 |
+
files = classification["files"] + segmentation["files"] + detection["files"] + [str(report_path)]
|
| 324 |
+
status = f"Complete multi-model assessment · {prepared.width}×{prepared.height} · {elapsed:.1f}s"
|
| 325 |
+
return (
|
| 326 |
+
summary,
|
| 327 |
+
classification["assessment"],
|
| 328 |
+
classification["rows"],
|
| 329 |
+
segmentation["overlay"],
|
| 330 |
+
segmentation["mask"],
|
| 331 |
+
segmentation["rows"],
|
| 332 |
+
detection["overlay"],
|
| 333 |
+
detection["summary"],
|
| 334 |
+
detection["details"],
|
| 335 |
+
files,
|
| 336 |
+
status,
|
| 337 |
)
|
|
|
|
| 338 |
|
| 339 |
|
| 340 |
CSS = """
|
| 341 |
+
.gradio-container {max-width: 1440px !important; background: #f6f8fb;}
|
| 342 |
+
.hero {padding: 2rem; border-radius: 22px; color: white; background: linear-gradient(125deg,#071c33,#0a4b5c 56%,#198f75); box-shadow: 0 18px 44px rgba(7,28,51,.18); margin-bottom: 1rem;}
|
| 343 |
+
.hero h1 {font-size: 2.35rem; margin: 0 0 .35rem; letter-spacing: -.03em;}
|
| 344 |
+
.hero p {max-width: 850px; margin: .35rem 0; color: #d8f3ee;}
|
| 345 |
+
.hero a {color: #fff; font-weight: 650;}
|
| 346 |
+
.pipeline {display:grid;grid-template-columns:repeat(3,1fr);gap:12px;margin:14px 0 20px;}
|
| 347 |
+
.pipeline div,.assessment-card,.metric-card {background:white;border:1px solid #dce6ed;border-radius:16px;padding:16px;box-shadow:0 6px 18px rgba(20,50,70,.06);}
|
| 348 |
+
.pipeline b {display:block;color:#0c5262;margin-bottom:5px}.pipeline span,.micro-note {color:#667985;font-size:.87rem;}
|
| 349 |
+
.summary-grid {display:grid;grid-template-columns:repeat(4,1fr);gap:12px;margin:12px 0;}
|
| 350 |
+
.metric-card span,.eyebrow {display:block;color:#66808c;font-size:.72rem;font-weight:750;letter-spacing:.1em;text-transform:uppercase;}
|
| 351 |
+
.metric-card strong {display:block;font-size:1.45rem;margin:6px 0;color:#113544;}.metric-card small {color:#60747e;}
|
| 352 |
+
.assessment-card h2 {margin:.25rem 0;color:#123c49}.assessment-card p {color:#526b76;}
|
| 353 |
+
.prob-row {display:grid;grid-template-columns:155px 1fr 62px;gap:10px;align-items:center;margin:8px 0;font-size:.86rem;}
|
| 354 |
+
.prob-row b {text-align:right}.prob-track {height:9px;background:#e5edf1;border-radius:20px;overflow:hidden}.prob-track i {display:block;height:100%;background:linear-gradient(90deg,#169c7d,#36b7c5);border-radius:20px;}
|
| 355 |
+
.section-note {padding:12px 14px;border-left:4px solid #15947a;background:#eef9f6;border-radius:8px;color:#315c62;}
|
| 356 |
+
@media(max-width:850px){.pipeline,.summary-grid{grid-template-columns:1fr}.prob-row{grid-template-columns:115px 1fr 56px}}
|
| 357 |
"""
|
| 358 |
|
| 359 |
+
|
| 360 |
+
with gr.Blocks(title="Satellite Vision Toolkit Pro", css=CSS, theme=gr.themes.Soft()) as demo:
|
| 361 |
gr.HTML("""
|
| 362 |
<div class="hero">
|
| 363 |
+
<div class="eyebrow" style="color:#8ee5d2">REMOTE SENSING DECISION SUPPORT</div>
|
| 364 |
+
<h1>🛰️ Satellite Vision Toolkit Pro</h1>
|
| 365 |
+
<p>A multi-level workbench for scene-level land-use/land-cover classification, pixel-level cover mapping, and overhead object detection.</p>
|
| 366 |
+
<p><a href="https://github.com/LabMingzeChen/SatelliteVisionToolkit">GitHub</a> · <a href="https://huggingface.co/mrm8488/convnext-tiny-finetuned-eurosat">LULC model</a> · <a href="https://huggingface.co/mfaytin/mask2former-satellite">Segmentation model</a> · <a href="https://huggingface.co/bluelabel/satellite-equipment-detection-yolov8n-vhr10">Detection model</a></p>
|
| 367 |
+
</div>
|
| 368 |
+
<div class="pipeline">
|
| 369 |
+
<div><b>01 · Scene classification</b><span>EuroSAT probability profile across 10 LULC scene types.</span></div>
|
| 370 |
+
<div><b>02 · Semantic segmentation</b><span>Per-pixel OpenEarthMap land-cover composition and masks.</span></div>
|
| 371 |
+
<div><b>03 · Object detection</b><span>Bounding boxes and inventory-style summaries for 10 VHR object types.</span></div>
|
| 372 |
</div>
|
| 373 |
""")
|
| 374 |
+
with gr.Row(equal_height=True):
|
| 375 |
+
image_input = gr.Image(type="pil", label="Satellite / aerial RGB image", height=430)
|
| 376 |
with gr.Column():
|
| 377 |
+
gr.Markdown("### Analysis controls\nTune reproducible thresholds, then run the complete assessment or an individual method.")
|
| 378 |
+
top_k = gr.Slider(3, 10, value=5, step=1, label="LULC alternatives (top-k)")
|
| 379 |
+
opacity = gr.Slider(0.1, 0.9, value=0.55, step=0.05, label="Segmentation overlay opacity")
|
| 380 |
+
min_share = gr.Slider(0.0, 5.0, value=0.1, step=0.1, label="Minimum reported cover share (%)")
|
| 381 |
+
confidence = gr.Slider(0.05, 0.9, value=0.25, step=0.05, label="Detection confidence threshold")
|
| 382 |
+
iou = gr.Slider(0.1, 0.9, value=0.45, step=0.05, label="Detection NMS IoU threshold")
|
| 383 |
+
analyze_button = gr.Button("Run complete professional assessment", variant="primary", size="lg")
|
| 384 |
|
| 385 |
+
with gr.Tabs():
|
| 386 |
+
with gr.Tab("Executive overview"):
|
| 387 |
+
analysis_status = gr.Markdown()
|
| 388 |
+
executive_summary = gr.HTML()
|
| 389 |
+
overview_lulc = gr.HTML()
|
| 390 |
+
overview_lulc_table = gr.Dataframe(
|
| 391 |
+
headers=["Rank", "LULC class", "Probability (%)", "Confidence tier"],
|
| 392 |
+
interactive=False,
|
| 393 |
+
label="Scene classification probability profile",
|
| 394 |
+
)
|
| 395 |
+
with gr.Row():
|
| 396 |
+
overview_segment = gr.Image(label="Pixel-level land-cover overlay")
|
| 397 |
+
overview_detection = gr.Image(label="Detected objects")
|
| 398 |
+
overview_files = gr.File(label="Download complete evidence package", file_count="multiple")
|
| 399 |
|
| 400 |
+
with gr.Tab("LULC classification"):
|
| 401 |
+
gr.Markdown("<div class='section-note'><b>Scene-level interpretation.</b> Assigns the whole image to EuroSAT land-use/land-cover classes. This is distinct from pixel segmentation and is not a legal land-use designation.</div>")
|
| 402 |
+
classify_button = gr.Button("Classify scene LULC", variant="primary")
|
| 403 |
+
classify_status = gr.Markdown()
|
| 404 |
+
classification_assessment = gr.HTML()
|
| 405 |
+
classification_table = gr.Dataframe(
|
| 406 |
+
headers=["Rank", "LULC class", "Probability (%)", "Confidence tier"],
|
| 407 |
+
interactive=False,
|
| 408 |
+
label="Ranked LULC alternatives",
|
| 409 |
+
)
|
| 410 |
+
classification_files = gr.File(label="Download classification CSV / JSON", file_count="multiple")
|
| 411 |
|
|
|
|
| 412 |
with gr.Tab("Land-cover segmentation"):
|
| 413 |
+
gr.Markdown("<div class='section-note'><b>Pixel-level interpretation.</b> Maps nine OpenEarthMap surface classes and reports image-pixel composition.</div>")
|
|
|
|
|
|
|
| 414 |
segment_button = gr.Button("Segment land cover", variant="primary")
|
| 415 |
segment_status = gr.Markdown()
|
| 416 |
with gr.Row():
|
|
|
|
| 418 |
segment_mask = gr.Image(label="Categorical mask")
|
| 419 |
segment_table = gr.Dataframe(
|
| 420 |
headers=["Class ID", "Class", "Pixels", "Share (%)", "Color"],
|
|
|
|
| 421 |
interactive=False,
|
| 422 |
label="Land-cover area summary",
|
| 423 |
)
|
| 424 |
segment_files = gr.File(label="Download segmentation outputs", file_count="multiple")
|
| 425 |
|
| 426 |
with gr.Tab("Object detection"):
|
| 427 |
+
gr.Markdown("<div class='section-note'><b>Instance-level interpretation.</b> Locates supported objects with confidence-filtered bounding boxes.</div>")
|
|
|
|
|
|
|
| 428 |
detect_button = gr.Button("Detect satellite objects", variant="primary")
|
| 429 |
detect_status = gr.Markdown()
|
| 430 |
detect_overlay = gr.Image(label="Detection overlay")
|
| 431 |
detection_summary = gr.Dataframe(
|
| 432 |
headers=["Class", "Count", "Average confidence", "Maximum confidence"],
|
|
|
|
| 433 |
interactive=False,
|
| 434 |
label="Detection summary",
|
| 435 |
)
|
|
|
|
| 440 |
)
|
| 441 |
detection_files = gr.File(label="Download detection outputs", file_count="multiple")
|
| 442 |
|
| 443 |
+
with gr.Tab("Methodology & scope"):
|
| 444 |
+
gr.Markdown("""
|
| 445 |
+
### Analytical hierarchy
|
| 446 |
+
|
| 447 |
+
| Level | Question answered | Model / training domain | Output |
|
| 448 |
+
|---|---|---|---|
|
| 449 |
+
| Scene | What broad LULC type best characterizes this image? | ConvNeXT-Tiny / EuroSAT Sentinel-2 RGB | Ranked probabilities + entropy |
|
| 450 |
+
| Pixel | Which cover class is predicted at each pixel? | Mask2Former / OpenEarthMap | Overlay, mask, pixel shares |
|
| 451 |
+
| Object | Where are supported discrete objects? | YOLOv8n / NWPU VHR-10 | Boxes, counts, CSV, pixel GeoJSON |
|
| 452 |
+
|
| 453 |
+
**Interpretation guardrails:** EuroSAT is a European Sentinel-2 scene dataset; classification may shift on other sensors, regions, resolutions, or crops. Pixel shares are not automatically physical ground-area shares. Pixel-coordinate GeoJSON is not georeferenced. Models can miss small or obscured objects. Do not use outputs alone for legal, surveillance, emergency, navigation, or safety-critical decisions.
|
| 454 |
+
""")
|
| 455 |
|
| 456 |
+
classify_button.click(
|
| 457 |
+
classify_lulc,
|
| 458 |
+
inputs=[image_input, top_k],
|
| 459 |
+
outputs=[classification_assessment, classification_table, classification_files, classify_status],
|
| 460 |
+
api_name="classify",
|
| 461 |
+
)
|
| 462 |
segment_button.click(
|
| 463 |
segment_satellite_image,
|
| 464 |
inputs=[image_input, opacity, min_share],
|
|
|
|
| 471 |
outputs=[detect_overlay, detection_summary, detection_details, detection_files, detect_status],
|
| 472 |
api_name="detect",
|
| 473 |
)
|
| 474 |
+
analyze_button.click(
|
| 475 |
+
analyze_satellite_image,
|
| 476 |
+
inputs=[image_input, top_k, opacity, min_share, confidence, iou],
|
| 477 |
+
outputs=[
|
| 478 |
+
executive_summary,
|
| 479 |
+
overview_lulc,
|
| 480 |
+
overview_lulc_table,
|
| 481 |
+
overview_segment,
|
| 482 |
+
segment_mask,
|
| 483 |
+
segment_table,
|
| 484 |
+
overview_detection,
|
| 485 |
+
detection_summary,
|
| 486 |
+
detection_details,
|
| 487 |
+
overview_files,
|
| 488 |
+
analysis_status,
|
| 489 |
+
],
|
| 490 |
+
api_name="analyze",
|
| 491 |
+
)
|
| 492 |
|
| 493 |
|
| 494 |
if __name__ == "__main__":
|
satellite_utils.py
CHANGED
|
@@ -3,7 +3,9 @@
|
|
| 3 |
from __future__ import annotations
|
| 4 |
|
| 5 |
import csv
|
|
|
|
| 6 |
import json
|
|
|
|
| 7 |
from collections import defaultdict
|
| 8 |
from pathlib import Path
|
| 9 |
from typing import Iterable
|
|
@@ -27,11 +29,119 @@ LAND_COVER_PALETTE: dict[str, tuple[int, int, int]] = {
|
|
| 27 |
"building": (218, 73, 73),
|
| 28 |
}
|
| 29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
def normalize_label(label: str) -> str:
|
| 32 |
return label.lower().replace("_", " ").strip()
|
| 33 |
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
def fallback_color(class_id: int) -> tuple[int, int, int]:
|
| 36 |
return (
|
| 37 |
int((67 * class_id + 41) % 190 + 35),
|
|
@@ -257,4 +367,3 @@ def write_pixel_geojson(
|
|
| 257 |
"features": features,
|
| 258 |
}
|
| 259 |
path.write_text(json.dumps(collection, indent=2), encoding="utf-8")
|
| 260 |
-
|
|
|
|
| 3 |
from __future__ import annotations
|
| 4 |
|
| 5 |
import csv
|
| 6 |
+
import html
|
| 7 |
import json
|
| 8 |
+
import math
|
| 9 |
from collections import defaultdict
|
| 10 |
from pathlib import Path
|
| 11 |
from typing import Iterable
|
|
|
|
| 29 |
"building": (218, 73, 73),
|
| 30 |
}
|
| 31 |
|
| 32 |
+
LULC_DISPLAY_NAMES = {
|
| 33 |
+
"annualcrop": "Annual crop",
|
| 34 |
+
"forest": "Forest",
|
| 35 |
+
"herbaceousvegetation": "Herbaceous vegetation",
|
| 36 |
+
"highway": "Highway",
|
| 37 |
+
"industrial": "Industrial",
|
| 38 |
+
"pasture": "Pasture",
|
| 39 |
+
"permanentcrop": "Permanent crop",
|
| 40 |
+
"residential": "Residential",
|
| 41 |
+
"river": "River",
|
| 42 |
+
"sealake": "Sea / lake",
|
| 43 |
+
}
|
| 44 |
+
|
| 45 |
|
| 46 |
def normalize_label(label: str) -> str:
|
| 47 |
return label.lower().replace("_", " ").strip()
|
| 48 |
|
| 49 |
|
| 50 |
+
def display_lulc_label(label: str) -> str:
|
| 51 |
+
"""Convert EuroSAT model labels into compact report labels."""
|
| 52 |
+
key = "".join(character for character in label.lower() if character.isalnum())
|
| 53 |
+
return LULC_DISPLAY_NAMES.get(key, label.replace("_", " ").strip().title())
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def confidence_tier(probability: float) -> str:
|
| 57 |
+
if probability >= 0.80:
|
| 58 |
+
return "High"
|
| 59 |
+
if probability >= 0.55:
|
| 60 |
+
return "Moderate"
|
| 61 |
+
return "Low"
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
def normalized_entropy(probabilities: Iterable[float]) -> float:
|
| 65 |
+
"""Return Shannon entropy normalized to 0–1 for model ambiguity."""
|
| 66 |
+
values = [max(0.0, float(value)) for value in probabilities]
|
| 67 |
+
total = sum(values)
|
| 68 |
+
if not values or total <= 0.0 or len(values) == 1:
|
| 69 |
+
return 0.0
|
| 70 |
+
normalized = [value / total for value in values if value > 0.0]
|
| 71 |
+
entropy = -sum(value * math.log(value) for value in normalized)
|
| 72 |
+
return entropy / math.log(len(values))
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def build_lulc_table(
|
| 76 |
+
probabilities: Iterable[float],
|
| 77 |
+
id2label: dict[int, str],
|
| 78 |
+
top_k: int = 5,
|
| 79 |
+
) -> list[list[object]]:
|
| 80 |
+
ranked = sorted(
|
| 81 |
+
enumerate(float(value) for value in probabilities),
|
| 82 |
+
key=lambda item: item[1],
|
| 83 |
+
reverse=True,
|
| 84 |
+
)[: max(1, int(top_k))]
|
| 85 |
+
return [
|
| 86 |
+
[rank, display_lulc_label(id2label.get(class_id, f"class_{class_id}")), round(score * 100, 2), confidence_tier(score)]
|
| 87 |
+
for rank, (class_id, score) in enumerate(ranked, start=1)
|
| 88 |
+
]
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
def render_lulc_assessment(rows: list[list[object]], entropy: float) -> str:
|
| 92 |
+
"""Render an accessible probability profile and uncertainty note."""
|
| 93 |
+
if not rows:
|
| 94 |
+
return "<div class='assessment-card'>No classification result.</div>"
|
| 95 |
+
top_probability = float(rows[0][2])
|
| 96 |
+
bars = "".join(
|
| 97 |
+
"<div class='prob-row'><span>{}</span><div class='prob-track'><i style='width:{:.2f}%'></i></div><b>{:.2f}%</b></div>".format(
|
| 98 |
+
html.escape(str(row[1])), float(row[2]), float(row[2])
|
| 99 |
+
)
|
| 100 |
+
for row in rows
|
| 101 |
+
)
|
| 102 |
+
ambiguity = "low" if entropy < 0.35 else "moderate" if entropy < 0.65 else "high"
|
| 103 |
+
return (
|
| 104 |
+
"<div class='assessment-card'>"
|
| 105 |
+
f"<div class='eyebrow'>SCENE-LEVEL LULC</div><h2>{html.escape(str(rows[0][1]))}</h2>"
|
| 106 |
+
f"<p><strong>{top_probability:.2f}%</strong> top-class confidence · "
|
| 107 |
+
f"{ambiguity} ambiguity (normalized entropy {entropy:.2f})</p>{bars}"
|
| 108 |
+
"<p class='micro-note'>A whole-scene EuroSAT label, not a cadastral or planning designation.</p></div>"
|
| 109 |
+
)
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def build_analysis_summary(
|
| 113 |
+
lulc_rows: list[list[object]],
|
| 114 |
+
entropy: float,
|
| 115 |
+
land_cover_rows: list[list[object]],
|
| 116 |
+
detection_rows: list[list[object]],
|
| 117 |
+
elapsed_seconds: float,
|
| 118 |
+
) -> str:
|
| 119 |
+
lulc_name = str(lulc_rows[0][1]) if lulc_rows else "Unavailable"
|
| 120 |
+
lulc_confidence = float(lulc_rows[0][2]) if lulc_rows else 0.0
|
| 121 |
+
cover_name = str(land_cover_rows[0][1]) if land_cover_rows else "Unavailable"
|
| 122 |
+
cover_share = float(land_cover_rows[0][3]) if land_cover_rows else 0.0
|
| 123 |
+
object_count = sum(int(row[1]) for row in detection_rows)
|
| 124 |
+
return f"""
|
| 125 |
+
<div class="summary-grid">
|
| 126 |
+
<div class="metric-card"><span>Scene LULC</span><strong>{html.escape(lulc_name)}</strong><small>{lulc_confidence:.1f}% confidence · entropy {entropy:.2f}</small></div>
|
| 127 |
+
<div class="metric-card"><span>Dominant cover</span><strong>{html.escape(cover_name)}</strong><small>{cover_share:.1f}% of processed pixels</small></div>
|
| 128 |
+
<div class="metric-card"><span>Detected objects</span><strong>{object_count}</strong><small>{len(detection_rows)} represented object classes</small></div>
|
| 129 |
+
<div class="metric-card"><span>Analysis time</span><strong>{elapsed_seconds:.1f}s</strong><small>classification + segmentation + detection</small></div>
|
| 130 |
+
</div>
|
| 131 |
+
"""
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
def write_lulc_csv(path: Path, rows: Iterable[Iterable[object]]) -> None:
|
| 135 |
+
with path.open("w", newline="", encoding="utf-8") as handle:
|
| 136 |
+
writer = csv.writer(handle)
|
| 137 |
+
writer.writerow(["rank", "class", "probability_percent", "confidence_tier"])
|
| 138 |
+
writer.writerows(rows)
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
def write_json(path: Path, payload: object) -> None:
|
| 142 |
+
path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")
|
| 143 |
+
|
| 144 |
+
|
| 145 |
def fallback_color(class_id: int) -> tuple[int, int, int]:
|
| 146 |
return (
|
| 147 |
int((67 * class_id + 41) % 190 + 35),
|
|
|
|
| 367 |
"features": features,
|
| 368 |
}
|
| 369 |
path.write_text(json.dumps(collection, indent=2), encoding="utf-8")
|
|
|
scripts/satellite_client.py
CHANGED
|
@@ -15,7 +15,7 @@ DEFAULT_SPACE = "Mingze/SatelliteVisionToolkit"
|
|
| 15 |
|
| 16 |
def parse_args() -> argparse.Namespace:
|
| 17 |
parser = argparse.ArgumentParser(description=__doc__)
|
| 18 |
-
parser.add_argument("operation", choices=("
|
| 19 |
parser.add_argument("image", type=Path)
|
| 20 |
parser.add_argument("--space", default=DEFAULT_SPACE)
|
| 21 |
parser.add_argument("--output", type=Path)
|
|
@@ -23,6 +23,7 @@ def parse_args() -> argparse.Namespace:
|
|
| 23 |
parser.add_argument("--iou", type=float, default=0.45)
|
| 24 |
parser.add_argument("--opacity", type=float, default=0.55)
|
| 25 |
parser.add_argument("--min-share", type=float, default=0.1)
|
|
|
|
| 26 |
return parser.parse_args()
|
| 27 |
|
| 28 |
|
|
@@ -31,20 +32,36 @@ def main() -> int:
|
|
| 31 |
if not args.image.is_file():
|
| 32 |
raise SystemExit(f"Image not found: {args.image}")
|
| 33 |
client = Client(args.space)
|
| 34 |
-
if args.operation == "
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
result = client.predict(
|
| 36 |
handle_file(str(args.image)),
|
| 37 |
args.confidence,
|
| 38 |
args.iou,
|
| 39 |
api_name="/detect",
|
| 40 |
)
|
| 41 |
-
|
| 42 |
result = client.predict(
|
| 43 |
handle_file(str(args.image)),
|
| 44 |
args.opacity,
|
| 45 |
args.min_share,
|
| 46 |
api_name="/segment",
|
| 47 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
rendered = json.dumps(result, ensure_ascii=False, indent=2, default=str)
|
| 49 |
if args.output:
|
| 50 |
args.output.write_text(rendered + "\n", encoding="utf-8")
|
|
@@ -55,4 +72,3 @@ def main() -> int:
|
|
| 55 |
|
| 56 |
if __name__ == "__main__":
|
| 57 |
raise SystemExit(main())
|
| 58 |
-
|
|
|
|
| 15 |
|
| 16 |
def parse_args() -> argparse.Namespace:
|
| 17 |
parser = argparse.ArgumentParser(description=__doc__)
|
| 18 |
+
parser.add_argument("operation", choices=("classify", "segment", "detect", "analyze"))
|
| 19 |
parser.add_argument("image", type=Path)
|
| 20 |
parser.add_argument("--space", default=DEFAULT_SPACE)
|
| 21 |
parser.add_argument("--output", type=Path)
|
|
|
|
| 23 |
parser.add_argument("--iou", type=float, default=0.45)
|
| 24 |
parser.add_argument("--opacity", type=float, default=0.55)
|
| 25 |
parser.add_argument("--min-share", type=float, default=0.1)
|
| 26 |
+
parser.add_argument("--top-k", type=int, default=5)
|
| 27 |
return parser.parse_args()
|
| 28 |
|
| 29 |
|
|
|
|
| 32 |
if not args.image.is_file():
|
| 33 |
raise SystemExit(f"Image not found: {args.image}")
|
| 34 |
client = Client(args.space)
|
| 35 |
+
if args.operation == "classify":
|
| 36 |
+
result = client.predict(
|
| 37 |
+
handle_file(str(args.image)),
|
| 38 |
+
args.top_k,
|
| 39 |
+
api_name="/classify",
|
| 40 |
+
)
|
| 41 |
+
elif args.operation == "detect":
|
| 42 |
result = client.predict(
|
| 43 |
handle_file(str(args.image)),
|
| 44 |
args.confidence,
|
| 45 |
args.iou,
|
| 46 |
api_name="/detect",
|
| 47 |
)
|
| 48 |
+
elif args.operation == "segment":
|
| 49 |
result = client.predict(
|
| 50 |
handle_file(str(args.image)),
|
| 51 |
args.opacity,
|
| 52 |
args.min_share,
|
| 53 |
api_name="/segment",
|
| 54 |
)
|
| 55 |
+
else:
|
| 56 |
+
result = client.predict(
|
| 57 |
+
handle_file(str(args.image)),
|
| 58 |
+
args.top_k,
|
| 59 |
+
args.opacity,
|
| 60 |
+
args.min_share,
|
| 61 |
+
args.confidence,
|
| 62 |
+
args.iou,
|
| 63 |
+
api_name="/analyze",
|
| 64 |
+
)
|
| 65 |
rendered = json.dumps(result, ensure_ascii=False, indent=2, default=str)
|
| 66 |
if args.output:
|
| 67 |
args.output.write_text(rendered + "\n", encoding="utf-8")
|
|
|
|
| 72 |
|
| 73 |
if __name__ == "__main__":
|
| 74 |
raise SystemExit(main())
|
|
|
skills/analyze-satellite-imagery/SKILL.md
CHANGED
|
@@ -1,37 +1,44 @@
|
|
| 1 |
---
|
| 2 |
name: analyze-satellite-imagery
|
| 3 |
-
description:
|
| 4 |
---
|
| 5 |
|
| 6 |
# Analyze Satellite Imagery
|
| 7 |
|
| 8 |
-
Use the bundled Space or local app to run
|
| 9 |
|
|
|
|
| 10 |
- Detect 10 NWPU VHR-10 object categories with the fine-tuned YOLOv8n model.
|
| 11 |
- Segment 9 OpenEarthMap land-cover categories with Mask2Former.
|
| 12 |
|
| 13 |
## Choose a workflow
|
| 14 |
|
| 15 |
-
1. Use
|
| 16 |
2. Use segmentation for per-pixel land-cover composition.
|
| 17 |
-
3.
|
| 18 |
-
4.
|
|
|
|
| 19 |
|
| 20 |
## Run the app
|
| 21 |
|
| 22 |
-
From the plugin root, install `requirements.txt` and run `python app.py`. For a deployed Space, use the browser UI or call
|
| 23 |
|
| 24 |
Use `scripts/satellite_client.py` for repeatable API calls:
|
| 25 |
|
| 26 |
```bash
|
|
|
|
| 27 |
python scripts/satellite_client.py detect image.jpg --output result.json
|
| 28 |
python scripts/satellite_client.py segment image.jpg --output result.json
|
|
|
|
| 29 |
```
|
| 30 |
|
| 31 |
Set `--space` when using a fork. The default is `Mingze/SatelliteVisionToolkit`.
|
| 32 |
|
| 33 |
## Interpret outputs
|
| 34 |
|
|
|
|
|
|
|
|
|
|
| 35 |
- Treat detection counts as visible-image estimates, not inventories.
|
| 36 |
- Treat class shares as proportions of processed image pixels, not physical land area.
|
| 37 |
- State that exported GeoJSON uses top-left-origin image pixels and has no geographic CRS.
|
|
@@ -41,5 +48,4 @@ Set `--space` when using a fork. The default is `Mingze/SatelliteVisionToolkit`.
|
|
| 41 |
|
| 42 |
## Report results
|
| 43 |
|
| 44 |
-
Include the model, thresholds, processed image dimensions, detected classes/counts or land-cover shares, and
|
| 45 |
-
|
|
|
|
| 1 |
---
|
| 2 |
name: analyze-satellite-imagery
|
| 3 |
+
description: Classify land use and land cover, segment surface classes, and detect remote-sensing objects in satellite or aerial RGB imagery. Use when Codex needs to analyze overhead PNG, JPEG, WebP, or TIFF images; assign EuroSAT LULC scene classes; locate airplanes, ships, vehicles, storage tanks, bridges, harbors, or sports facilities; map OpenEarthMap land-cover classes; produce professional summaries, overlays, masks, CSV, JSON, or pixel-coordinate GeoJSON; or call the Satellite Vision Toolkit Hugging Face Space API.
|
| 4 |
---
|
| 5 |
|
| 6 |
# Analyze Satellite Imagery
|
| 7 |
|
| 8 |
+
Use the bundled Space or local app to run three complementary analytical levels:
|
| 9 |
|
| 10 |
+
- Classify the whole scene across 10 EuroSAT LULC categories with ConvNeXT-Tiny.
|
| 11 |
- Detect 10 NWPU VHR-10 object categories with the fine-tuned YOLOv8n model.
|
| 12 |
- Segment 9 OpenEarthMap land-cover categories with Mask2Former.
|
| 13 |
|
| 14 |
## Choose a workflow
|
| 15 |
|
| 16 |
+
1. Use classification for a broad whole-scene LULC hypothesis and ranked alternatives.
|
| 17 |
2. Use segmentation for per-pixel land-cover composition.
|
| 18 |
+
3. Use detection for discrete objects and bounding boxes.
|
| 19 |
+
4. Use complete analysis when a professional summary or cross-method evidence package is needed.
|
| 20 |
+
5. Inspect `references/model-guide.md` before making claims about model scope, licenses, or limitations.
|
| 21 |
|
| 22 |
## Run the app
|
| 23 |
|
| 24 |
+
From the plugin root, install `requirements.txt` and run `python app.py`. For a deployed Space, use the browser UI or call `/classify`, `/segment`, `/detect`, or `/analyze` with `gradio_client`.
|
| 25 |
|
| 26 |
Use `scripts/satellite_client.py` for repeatable API calls:
|
| 27 |
|
| 28 |
```bash
|
| 29 |
+
python scripts/satellite_client.py classify image.jpg --output classification.json
|
| 30 |
python scripts/satellite_client.py detect image.jpg --output result.json
|
| 31 |
python scripts/satellite_client.py segment image.jpg --output result.json
|
| 32 |
+
python scripts/satellite_client.py analyze image.jpg --output complete.json
|
| 33 |
```
|
| 34 |
|
| 35 |
Set `--space` when using a fork. The default is `Mingze/SatelliteVisionToolkit`.
|
| 36 |
|
| 37 |
## Interpret outputs
|
| 38 |
|
| 39 |
+
- Keep scene-level classification separate from pixel-level segmentation in the report.
|
| 40 |
+
- Treat EuroSAT probabilities as a broad scene hypothesis, not zoning, cadastral, or legal land-use evidence.
|
| 41 |
+
- Report normalized entropy when classification ambiguity matters; a high top score does not remove domain-shift risk.
|
| 42 |
- Treat detection counts as visible-image estimates, not inventories.
|
| 43 |
- Treat class shares as proportions of processed image pixels, not physical land area.
|
| 44 |
- State that exported GeoJSON uses top-left-origin image pixels and has no geographic CRS.
|
|
|
|
| 48 |
|
| 49 |
## Report results
|
| 50 |
|
| 51 |
+
Include the analytical level, model, thresholds, processed image dimensions, ranked LULC alternatives, detected classes/counts or land-cover shares, and important limitations. Link or return generated overlays and exports when available.
|
|
|
skills/analyze-satellite-imagery/agents/openai.yaml
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
interface:
|
| 2 |
display_name: "Satellite Imagery Analysis"
|
| 3 |
-
short_description: "
|
| 4 |
-
default_prompt: "Use $analyze-satellite-imagery to
|
|
|
|
| 1 |
interface:
|
| 2 |
display_name: "Satellite Imagery Analysis"
|
| 3 |
+
short_description: "Classify LULC, segment cover, and detect overhead objects"
|
| 4 |
+
default_prompt: "Use $analyze-satellite-imagery to run a professional scene, pixel, and object-level assessment of this satellite image."
|
skills/analyze-satellite-imagery/references/model-guide.md
CHANGED
|
@@ -1,5 +1,16 @@
|
|
| 1 |
# Model and output guide
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
## Object detection
|
| 4 |
|
| 5 |
- Model: `bluelabel/satellite-equipment-detection-yolov8n-vhr10`
|
|
@@ -24,8 +35,9 @@ Area share is computed as `class pixels / all processed pixels`. It is a two-dim
|
|
| 24 |
|
| 25 |
## Export semantics
|
| 26 |
|
|
|
|
|
|
|
| 27 |
- Segmentation class-ID PNG preserves numeric predicted labels.
|
| 28 |
- Segmentation CSV contains class ID, name, pixel count, share percent, and display color.
|
| 29 |
- Detection CSV contains image-pixel boxes, pixel area, and normalized box centers.
|
| 30 |
- Detection GeoJSON stores box polygons in image-pixel coordinates with origin at the top left. It is intentionally not assigned a geographic coordinate reference system.
|
| 31 |
-
|
|
|
|
| 1 |
# Model and output guide
|
| 2 |
|
| 3 |
+
## Scene-level LULC classification
|
| 4 |
+
|
| 5 |
+
- Model: `mrm8488/convnext-tiny-finetuned-eurosat`
|
| 6 |
+
- Architecture: ConvNeXT-Tiny
|
| 7 |
+
- Training data: EuroSAT RGB images derived from European Sentinel-2 imagery
|
| 8 |
+
- Classes: annual crop, forest, herbaceous vegetation, highway, industrial, pasture, permanent crop, residential, river, sea/lake
|
| 9 |
+
- Model-card license: Apache-2.0
|
| 10 |
+
- Model-card reported evaluation accuracy: 0.9805
|
| 11 |
+
|
| 12 |
+
The classifier assigns one broad label to the processed image. It does not delineate parcels or pixels and must not be represented as a cadastral, zoning, or legal land-use conclusion. EuroSAT images are small European Sentinel-2 tiles; other sensors, countries, seasons, scales, and image crops introduce domain shift. Review the full probability profile and normalized entropy, not only the top label.
|
| 13 |
+
|
| 14 |
## Object detection
|
| 15 |
|
| 16 |
- Model: `bluelabel/satellite-equipment-detection-yolov8n-vhr10`
|
|
|
|
| 35 |
|
| 36 |
## Export semantics
|
| 37 |
|
| 38 |
+
- Classification CSV and JSON preserve ranked probabilities, confidence tiers, and normalized entropy.
|
| 39 |
+
- Complete-analysis JSON records all three model IDs, summaries, thresholds-derived outputs, timing, and coordinate caveats.
|
| 40 |
- Segmentation class-ID PNG preserves numeric predicted labels.
|
| 41 |
- Segmentation CSV contains class ID, name, pixel count, share percent, and display color.
|
| 42 |
- Detection CSV contains image-pixel boxes, pixel area, and normalized box centers.
|
| 43 |
- Detection GeoJSON stores box polygons in image-pixel coordinates with origin at the top left. It is intentionally not assigned a geographic coordinate reference system.
|
|
|
tests/test_app_contract.py
CHANGED
|
@@ -2,7 +2,7 @@ import ast
|
|
| 2 |
from pathlib import Path
|
| 3 |
|
| 4 |
|
| 5 |
-
def
|
| 6 |
source = Path("app.py").read_text(encoding="utf-8")
|
| 7 |
tree = ast.parse(source)
|
| 8 |
constants = {
|
|
@@ -12,6 +12,9 @@ def test_app_exposes_both_api_endpoints():
|
|
| 12 |
}
|
| 13 |
assert "segment" in constants
|
| 14 |
assert "detect" in constants
|
|
|
|
|
|
|
|
|
|
| 15 |
assert "mfaytin/mask2former-satellite" in constants
|
| 16 |
assert "bluelabel/satellite-equipment-detection-yolov8n-vhr10" in constants
|
| 17 |
-
|
|
|
|
| 2 |
from pathlib import Path
|
| 3 |
|
| 4 |
|
| 5 |
+
def test_app_exposes_all_api_endpoints_and_models():
|
| 6 |
source = Path("app.py").read_text(encoding="utf-8")
|
| 7 |
tree = ast.parse(source)
|
| 8 |
constants = {
|
|
|
|
| 12 |
}
|
| 13 |
assert "segment" in constants
|
| 14 |
assert "detect" in constants
|
| 15 |
+
assert "classify" in constants
|
| 16 |
+
assert "analyze" in constants
|
| 17 |
+
assert "mrm8488/convnext-tiny-finetuned-eurosat" in constants
|
| 18 |
assert "mfaytin/mask2former-satellite" in constants
|
| 19 |
assert "bluelabel/satellite-equipment-detection-yolov8n-vhr10" in constants
|
| 20 |
+
assert "cropland" in constants
|
tests/test_satellite_utils.py
CHANGED
|
@@ -52,6 +52,23 @@ def test_class_table_is_sorted_and_thresholded():
|
|
| 52 |
assert rows == [[4, "road", 3, 75.0, "#5C5C5C"]]
|
| 53 |
|
| 54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
def test_segmentation_outputs_match_input_size():
|
| 56 |
image = Image.new("RGB", (3, 2), "black")
|
| 57 |
class_map = np.array([[4, 4, 6], [4, 6, 6]], dtype=np.uint8)
|
|
@@ -81,4 +98,3 @@ def test_pixel_geojson_is_explicitly_unreferenced(tmp_path):
|
|
| 81 |
assert data["properties"]["coordinate_system"] == "image_pixels"
|
| 82 |
assert data["properties"]["origin"] == "top_left"
|
| 83 |
assert data["features"][0]["geometry"]["coordinates"][0][0] == [10.0, 20.0]
|
| 84 |
-
|
|
|
|
| 52 |
assert rows == [[4, "road", 3, 75.0, "#5C5C5C"]]
|
| 53 |
|
| 54 |
|
| 55 |
+
def test_lulc_table_is_ranked_and_human_readable():
|
| 56 |
+
rows = utils.build_lulc_table(
|
| 57 |
+
[0.1, 0.65, 0.25],
|
| 58 |
+
{0: "AnnualCrop", 1: "SeaLake", 2: "HerbaceousVegetation"},
|
| 59 |
+
top_k=2,
|
| 60 |
+
)
|
| 61 |
+
assert rows == [
|
| 62 |
+
[1, "Sea / lake", 65.0, "Moderate"],
|
| 63 |
+
[2, "Herbaceous vegetation", 25.0, "Low"],
|
| 64 |
+
]
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def test_normalized_entropy_has_expected_extremes():
|
| 68 |
+
assert utils.normalized_entropy([1.0, 0.0, 0.0]) == 0.0
|
| 69 |
+
assert round(utils.normalized_entropy([1 / 3, 1 / 3, 1 / 3]), 6) == 1.0
|
| 70 |
+
|
| 71 |
+
|
| 72 |
def test_segmentation_outputs_match_input_size():
|
| 73 |
image = Image.new("RGB", (3, 2), "black")
|
| 74 |
class_map = np.array([[4, 4, 6], [4, 6, 6]], dtype=np.uint8)
|
|
|
|
| 98 |
assert data["properties"]["coordinate_system"] == "image_pixels"
|
| 99 |
assert data["properties"]["origin"] == "top_left"
|
| 100 |
assert data["features"][0]["geometry"]["coordinates"][0][0] == [10.0, 20.0]
|
|
|