RESISC45-SigLIP2 is a vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for multi-label image classification. It is specifically trained to recognize and tag multiple land use and land cover scene categories from the RESISC45 dataset using the SiglipForImageClassification architecture.
[!note] SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features https://arxiv.org/pdf/2502.14786
1Classification Report:
2 precision recall f1-score support
3
4 airplane 0.9830 0.9900 0.9865 700
5 airport 0.9461 0.9529 0.9495 700
6 baseball diamond 0.9802 0.9886 0.9844 700
7 basketball court 0.9516 0.9271 0.9392 700
8 beach 0.9914 0.9900 0.9907 700
9 bridge 0.9730 0.9771 0.9751 700
10 chaparral 0.9957 0.9986 0.9971 700
11 church 0.7949 0.8971 0.8430 700
12 circular farmland 0.9914 0.9914 0.9914 700
13 cloud 0.9957 0.9871 0.9914 700
14 commercial area 0.9231 0.8229 0.8701 700
15 dense residential 0.9355 0.8914 0.9129 700
16 desert 0.9821 0.9414 0.9613 700
17 forest 0.9652 0.9514 0.9583 700
18 freeway 0.9344 0.9571 0.9457 700
19 golf course 0.9759 0.9843 0.9801 700
20 ground track field 0.9623 0.9857 0.9739 700
21 harbor 0.9885 0.9843 0.9864 700
22 industrial area 0.9505 0.9043 0.9268 700
23 intersection 0.9855 0.9686 0.9769 700
24 island 0.9871 0.9829 0.9850 700
25 lake 0.9440 0.9629 0.9533 700
26 meadow 0.9564 0.9400 0.9481 700
27 medium residential 0.8602 0.9314 0.8944 700
28 mobile home park 0.9610 0.9500 0.9555 700
29 mountain 0.9388 0.9429 0.9408 700
30 overpass 0.9614 0.9614 0.9614 700
31 palace 0.8455 0.8286 0.8369 700
32 parking lot 0.9899 0.9757 0.9827 700
33 railway 0.9407 0.9071 0.9236 700
34 railway station 0.9104 0.9143 0.9123 700
35 rectangular farmland 0.9572 0.9271 0.9419 700
36 river 0.9281 0.9586 0.9431 700
37 roundabout 0.9914 0.9871 0.9893 700
38 runway 0.9669 0.9586 0.9627 700
39 sea ice 0.9957 0.9943 0.9950 700
40 ship 0.9558 0.9886 0.9719 700
41 snowberg 0.9886 0.9900 0.9893 700
42 sparse residential 0.9238 0.9700 0.9463 700
43 stadium 0.9716 0.9757 0.9736 700
44 storage tank 0.9787 0.9829 0.9808 700
45 tennis court 0.9326 0.9486 0.9405 700
46 terrace 0.9372 0.9586 0.9477 700
47thermal power station 0.9482 0.9671 0.9576 700
48 wetland 0.9444 0.8986 0.9209 700
49
50 accuracy 0.9532 31500
51 macro avg 0.9538 0.9532 0.9532 31500
52 weighted avg 0.9538 0.9532 0.9532 31500Class 0: "airplane"
Class 1: "airport"
Class 2: "baseball diamond"
Class 3: "basketball court"
Class 4: "beach"
Class 5: "bridge"
Class 6: "chaparral"
Class 7: "church"
Class 8: "circular farmland"
Class 9: "cloud"
Class 10: "commercial area"
Class 11: "dense residential"
Class 12: "desert"
Class 13: "forest"
Class 14: "freeway"
Class 15: "golf course"
Class 16: "ground track field"
Class 17: "harbor"
Class 18: "industrial area"
Class 19: "intersection"
Class 20: "island"
Class 21: "lake"
Class 22: "meadow"
Class 23: "medium residential"
Class 24: "mobile home park"
Class 25: "mountain"
Class 26: "overpass"
Class 27: "palace"
Class 28: "parking lot"
Class 29: "railway"
Class 30: "railway station"
Class 31: "rectangular farmland"
Class 32: "river"
Class 33: "roundabout"
Class 34: "runway"
Class 35: "sea ice"
Class 36: "ship"
Class 37: "snowberg"
Class 38: "sparse residential"
Class 39: "stadium"
Class 40: "storage tank"
Class 41: "tennis court"
Class 42: "terrace"
Class 43: "thermal power station"
Class 44: "wetland"pip install -q transformers torch pillow gradio1import gradio as gr
2from transformers import AutoImageProcessor, SiglipForImageClassification
3from PIL import Image
4import torch
5
6# Load model and processor
7model_name = "prithivMLmods/RESISC45-SigLIP2" # Update to your actual Hugging Face model path
8model = SiglipForImageClassification.from_pretrained(model_name)
9processor = AutoImageProcessor.from_pretrained(model_name)
10
11# Label map
12id2label = {
13 "0": "airplane", "1": "airport", "2": "baseball diamond", "3": "basketball court", "4": "beach",
14 "5": "bridge", "6": "chaparral", "7": "church", "8": "circular farmland", "9": "cloud",
15 "10": "commercial area", "11": "dense residential", "12": "desert", "13": "forest", "14": "freeway",
16 "15": "golf course", "16": "ground track field", "17": "harbor", "18": "industrial area", "19": "intersection",
17 "20": "island", "21": "lake", "22": "meadow", "23": "medium residential", "24": "mobile home park",
18 "25": "mountain", "26": "overpass", "27": "palace", "28": "parking lot", "29": "railway",
19 "30": "railway station", "31": "rectangular farmland", "32": "river", "33": "roundabout", "34": "runway",
20 "35": "sea ice", "36": "ship", "37": "snowberg", "38": "sparse residential", "39": "stadium",
21 "40": "storage tank", "41": "tennis court", "42": "terrace", "43": "thermal power station", "44": "wetland"
22}
23
24def classify_resisc_image(image):
25 image = Image.fromarray(image).convert("RGB")
26 inputs = processor(images=image, return_tensors="pt")
27
28 with torch.no_grad():
29 outputs = model(**inputs)
30 logits = outputs.logits
31 probs = torch.sigmoid(logits).squeeze().tolist()
32
33 threshold = 0.5
34 predictions = {
35 id2label[str(i)]: round(probs[i], 3)
36 for i in range(len(probs)) if probs[i] >= threshold
37 }
38
39 return predictions or {"None Detected": 0.0}
40
41# Gradio Interface
42iface = gr.Interface(
43 fn=classify_resisc_image,
44 inputs=gr.Image(type="numpy"),
45 outputs=gr.Label(label="Predicted Scene Categories"),
46 title="RESISC45-SigLIP2",
47 description="Upload a satellite image to detect multiple land use and land cover categories (e.g., airport, forest, mountain)."
48)
49
50if __name__ == "__main__":
51 iface.launch()