This is a fine-tuned version of LiquidAI's LFM2.5-VL-450M vision-language model, specialized for satellite terrain classification. The model was trained in two stages:

VRSBench Training: Base training on VRSBench dataset
EuroSAT Fine-tuning: Additional training on EuroSAT land cover classification dataset

The model can classify satellite images into 10 land cover classes: AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake.

Training Details

Stage 1: VRSBench Pre-training

Base Model: LFM2.5-VL-450M
Dataset: VRSBench
Epochs: 1
Method: LoRA (r=16, alpha=32)

Stage 2: EuroSAT Fine-tuning

Base Model: VRSBench-trained model
Dataset: EuroSAT (27,000 satellite images, 64x64 RGB)
Training Samples: 21,600
Epochs: 2
Method: LoRA (r=16, alpha=32)
Hardware: Local training (no Ray/distributed)

Evaluation Results

EuroSAT Test Set (5,400 images)

Model	Accuracy
Base VRSBench Model	~10% (random baseline)
VRSBench + EuroSAT (this model)	97.52%

The model achieves near-perfect classification accuracy on EuroSAT, demonstrating significant improvement over the base VRSBench model.

Usage

With llama.cpp

bash

# Download Q4_K_M quantized version (recommended)
wget https://huggingface.co/5ch4um1/lfm2.5-vrsbench-EUROSAT-terrain-lora-450m/resolve/main/lfm2.5-vrsbench-terrain-expert-450m-q4_k_m.gguf

# Run inference
./llama-cli -m lfm2.5-vrsbench-terrain-expert-450m-q4_k_m.gguf \
  --image satellite_image.jpg \
  -p "What type of terrain is shown in this satellite image? Choose from: AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake."

With Transformers

python

from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image

model = AutoModelForVision2Seq.from_pretrained(
    "5ch4um1/lfm2.5-vrsbench-EUROSAT-terrain-lora-450m",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("5ch4um1/lfm2.5-vrsbench-EUROSAT-terrain-lora-450m")

image = Image.open("satellite_image.jpg")
prompt = "What type of terrain is shown in this satellite image? Choose from: AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake."

inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=20)
print(processor.decode(outputs[0], skip_special_tokens=True))

GGUF Quantizations

Version	Size	Description
F16	679 MB	Full precision (16-bit)
Q8_0	362 MB	8-bit quantization
Q4_K_M	219 MB	4-bit quantization (recommended for most use cases)

Model Sources

Base Model: LiquidAI/LFM2.5-VL-450M
EuroSAT Dataset: EuroSAT Paper

Limitations

The model is specialized for EuroSAT land cover classes and may not generalize to other satellite image classification tasks without additional training.
Images should be similar to EuroSAT format (RGB, overhead satellite view).
The model works best with 64x64 pixel images as used in training.

Training Environment

Framework: Transformers + PEFT (LoRA)
Hardware: Local GPU (CUDA)
Training Scripts: Available in the cookbook repository