XinRan-Tang/MMDiff-WHU-dataset

MMDiff: Multi-modal Remote Sensing Image Generation via Cross-Modality Spatial Feature Transfer

ISPRS 2026 🔥

Haojun Tang¹ · Wenda Zhao^1,* · Hengshuai Cui¹ · Haipeng Wang²

¹ Dalian University of Technology
² Unit 92728 of PLA

^* Corresponding author:

Abstract

Collecting spatially consistent multi-modal remote sensing (MMRS) images remains challenging due to different sensors vary in the imaging principles and acquisition times. This hinders the development of data-driven MMRS technologies, which rely on large-scale training samples. This paper proposes MMDiff, the first text-driven diffusion framework explicitly designed for jointly generating structurally consistent optical (OPT), synthetic aperture radar (SAR), and infrared (IR) remote sensing images from a single text prompt via cross-modality spatial feature transfer. MMDiff first trains the OPT branch with paired optical image-text data to capture rich semantic content, and then trains the SAR/IR branches with simple modality-specific text templates to learn the corresponding style attributes, without relying on complex linguistic descriptions. Specifically, we introduce a LoRA-based modality translation adaptation mechanism to translate the style attributes of optical spatial representations to SAR and IR style attributes while preserving the underlying semantic content. The translated representations are then transferred into the SAR and IR generation branches through the proposed spatial feature transfer mechanism, enabling rich spatial details in the generated SAR/IR images while maintaining cross-modal spatial consistency. Extensive experiments demonstrate that MMDiff achieves superior image quality in terms of modality similarity and semantic consistency compared to the state-of-the-art methods. Furthermore, MMDiff benefits downstream data-driven MMRS applications, e.g., multi-modal image fusion and object classification.

For each group, from left to right: optical (OPT), synthetic aperture radar (SAR), and near-infrared (NIR) images.

Dataset Overview

MMDiff-WHU-dataset is a multi-modal remote sensing image-text dataset derived from the WHU-OPT-SAR dataset. The WHU-OPT-SAR imagery was first cropped into 512×512 tiles, and each tile was then assigned to the semantic class occupying the largest area in its corresponding segmentation mask (i.e., a majority vote by area). The original WHU-OPT-SAR dataset contains seven land-cover classes. We retain four classes—city, farmland, forest, and water—by merging city and village into a single category and excluding road, others, and background. Finally, every tile was captioned with GPT-4o-mini, yielding co-registered optical (OPT), synthetic aperture radar (SAR), and near-infrared (NIR) image-text pairs.

Dataset Statistics

The dataset contains 20,943 image-text pairs across three modalities and occupies approximately 1.2 GB.

Modality	Training pairs	Test pairs	Total
Optical (OPT)	5,584	1,397	6,981
Synthetic aperture radar (SAR)	5,584	1,397	6,981
Near-infrared (NIR)	5,584	1,397	6,981

The three modalities share the same tiles, so the class distribution is identical across modalities:

Class	Training	Test	Total
city	280	70	350
farmland	2,392	598	2,990
forest	2,327	582	2,909
water	585	147	732
Total	5,584	1,397	6,981

Directory Structure

MMDiff-WHU-dataset/
├── opt/
│   ├── train/{city,farmland,forest,water}/*.jpg
│   ├── test/{city,farmland,forest,water}/*.jpg
│   ├── train.jsonl
│   └── test.jsonl
├── sar/
│   ├── train/{city,farmland,forest,water}/*.jpg
│   ├── test/{city,farmland,forest,water}/*.jpg
│   ├── train.jsonl
│   └── test.jsonl
└── ir/
    ├── train/{city,farmland,forest,water}/*.jpg
    ├── test/{city,farmland,forest,water}/*.jpg
    ├── train.jsonl
    └── test.jsonl

Each JSONL record contains a modality-relative image path and its text description:

{"image_path": "train/city/NH49E001014_0_2560_RGB.jpg", "text": "a satellite image of a city with a runway"}

Loading the Annotations

import json
from pathlib import Path

root = Path("/path/to/MMDiff-WHU-dataset")
with (root / "opt" / "train.jsonl").open() as f:
    samples = [json.loads(line) for line in f]

image_path = root / "opt" / samples[0]["image_path"]
caption = samples[0]["text"]

Intended Uses

The dataset can be used for modality-specific diffusion training, text-to-image generation, cross-modal spatial consistency evaluation, multi-modal image fusion, and downstream remote sensing land-cover recognition tasks. Please keep the original modality and split definitions when comparing results.