EarthJev

EarthJev

Answers typed questions about satellite image patches from all 12 Sentinel-2 bands, returning a probability for each option without writing text. A Qwen3.5-0.8B fine-tune for Earth observation researchers.

Made by dBenf on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: EarthJev is a small vision language model for Earth observation, from a team at the University of Naples Federico II. You give it a Sentinel-2 L2A patch with all 12 bands and a question with fixed options. It returns a probability for every option and never generates free text, so each answer is one of the options you declared.

Who it suits: Researchers and developers who classify satellite imagery and want fast, calibrated answers to fixed questions about a patch.

Industry | Who uses it | How they use it | Example
Earth observation | Remote sensing researchers | Ask land cover questions about Sentinel-2 patches | Asking which land cover class covers most of a patch
Environmental monitoring | Analysts who track land use | Check for features such as urban fabric across many patches | Asking whether a patch contains any urban area
Machine learning research | Developers who study efficient models | Compare its speed and accuracy with larger generative models | Timing answers per image against a generative baseline

Skip it if you need free text descriptions, or if your imagery has only the three colour bands.

Where to use it: It runs in Python with transformers and torch, from a clone of the project repository. The files download on Windows, macOS or Linux.

• Python with transformers that supports Qwen3.5, and torch
• The engine from the OneJev project
• Sentinel-2 L2A data in the BigEarthNet style

Use case | Who | What they get back
Land cover questions | Researchers | A probability for each land cover class
Yes or no checks | Analysts | A probability for yes and for no
Batch question answering | Developers | Several answers about one patch from one image encoding

How to set it up: You need Python, a GPU is not named as a requirement, and the model folder.

• Python with pip
• transformers with Qwen3.5 support, and torch
• The OneJev engine, whose install command is on the card
• A clone of this repository, which holds the earthjev_infer module

1. Install the Hugging Face hub library.
2. Download the model files.
3. Install transformers, torch and the OneJev engine.
4. Run the example from the clone of the repository.

pip install -U huggingface_hub
hf download dBenf/EarthJev

How to use it: Load the model from its folder, pass a 12 band array and a set of typed questions, and read the probabilities. The array stacks the 12 Sentinel-2 L2A bands, from B01 up to B12, along its first axis.

import numpy as np
from earthjev_infer import EarthJev

model = EarthJev.load(".")                       # this folder (downloaded with huggingface_hub.snapshot_download)
s2 = np.load("examples/S2A_MSIL2A_20170701T093031_N9999_R136_T35VPK_37_55.npy")   # [12, 120, 120] uint16
answers = model.predict(s2, {
    "dominant": {"type": "choice", "instructions": "Which land-cover class covers most of this patch?",
                 "criteria": {"artificial": "Artificial surfaces.", "agricultural": "Agricultural areas.",
                              "forest": "Forest and seminatural areas.", "wetlands": "Wetlands.",
                              "water": "Water bodies."}},
    "urban": {"type": "noul", "instructions": "Is there any urban fabric in this patch?"},
})
print(answers["dominant"])   # a probability for every option

Tool | What it does
EarthJev.load | Loads the model from the downloaded folder
model.predict | Takes the band array and typed questions, and returns a probability per option

Example: You pass a 12 band patch and ask which land cover class covers most of it, with five options. The model returns a probability for each of the five, such as one for forest and one for water, and no generated text.

For newcomers and experts: If you are new to this, Sentinel-2 is a satellite that images the Earth in many bands. Start with the example patch in the repository and read the probabilities for the two sample questions.

If you know the field, the model reads the logits of the option letters at one position, so every answer is valid and questions on one patch share one image encoding. The first vision layer was widened from 3 to 12 channels. The authors report calibration temperatures fitted on validation questions.

12 band patch and typed question
One image encoding is shared
Option letter logits are read
A probability per option returns

Good to know: Of three training runs, the authors kept the one with the best validation accuracy, 85.63 percent over 36,698 questions. In the card's tests it was several times faster than generative models of similar accuracy.

Title: Card test on CORINE questions
Model | Accuracy percent | Milliseconds per question
EarthJev | 86.0 | 19.9
Qwen3.5-0.8B, generative | 86.8 | 134.7
InternVL3-2B, LoRA | 87.1 | 84.5

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the Sentinel-2 patches you pass in
• Not for: free text answers, or imagery with only the three colour bands, since replacing the other nine bands with their training mean costs 4.5 points on CORINE

On Hugging Face, 14 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 1.1B
• Weights: about 2.2 GB in BF16; about 555 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 159
• Likes: 2
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-0.8B
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, remote-sensing, sentinel-2, multispectral, earth-observation, visual-question-answering, calibration, decision-model, system-one

Links

• Model: https://huggingface.co/dBenf/EarthJev
• Files: https://huggingface.co/dBenf/EarthJev/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download dBenf/EarthJev

Then follow the usage in its model card. Nothing in this kit runs by itself.