TAMMI — Text and Multi-Modal Imagery

TAMMI is the public release of TAMMI (Text and Multi-Modal Imagery), a large-scale multimodal Visual Question Answering (VQA) dataset for remote sensing, introduced at CVPR 2025 EarthVision Workshop.

TAMMI combines three complementary image modalities — Very High Resolution (VHR) orthophotos, Sentinel-2 multispectral imagery, and Sentinel-1 SAR — with rich question-answer annotations automatically generated from official geographic databases. It covers 6 French departments spanning urban, mountainous, and coastal landscapes.

---

Paper

Visual Question Answering on Multiple Remote Sensing Image Modalities
Hichem Boussaid*, Lucrezia Tosato*, Flora Weissgerber, Camille Kurtz, Laurent Wendling, Sylvain Lobry
LIPADE, Université Paris Cité & ONERA, France
CVPR 2025 EarthVision Workshop
Paper | Project Page

---

Dataset Summary

Split	Unique Images	Q&A Pairs	Avg Q&A / Image
Train	169,582	1,895,221	~11.2
Val	56,524	634,135	~11.2
Test	22,266	237,465	~10.7
Total	248,372	2,766,821

Each row corresponds to one unique image patch with all associated Q&A pairs grouped together (up to 50 pairs per image).

Note on test split: The test set is a curated subset of the full test partition (56,572 images), representing ~40% of images. The full dataset described in the paper (282,852 images) includes the complete test partition.

---

Geographic Coverage

Department	Region	Landscape
75 (Paris)	Île-de-France	Urban
92 (Hauts-de-Seine)	Île-de-France	Urban
93 (Seine-Saint-Denis)	Île-de-France	Urban
94 (Val-de-Marne)	Île-de-France	Urban
74 (Haute-Savoie)	Auvergne-Rhône-Alpes	Mountainous
34 (Hérault)	Occitanie	Maritime

Note: Departments 92, 93, and 94 share the same S1/S2 mosaics as department 75.

---

Image Modalities

Modality	Source	Resolution	Acquisition
BDOrtho (VHR RGB)	IGN	20 cm	2020 (dept. 74), 2021 (others)
Sentinel-2 (MS)	ESA Copernicus	10–20 m	L1C, cloud cover < 3%
Sentinel-1 (SAR)	ESA Copernicus	5×20 m (IW SLC)	VV and VH polarizations

VHR patches are 1000×1000 pixels (200m × 200m). MS and SAR patches are centered on each VHR patch and sized dynamically to provide spatial context.

---

Dataset Structure

Each row contains:

{
    "image_id":         str,       # unique identifier, e.g. "75-1004"
    "department":       str,       # French department code, e.g. "75"
    "patch_id":         str,       # patch identifier within department
    "bdo":              Image,     # VHR RGB orthophoto (1000×1000 px)
    "path_bdo":         str,       # relative path to BDO file
    "path_s1":          str,       # relative path to S1 patch
    "path_s2":          str,       # relative path to S2 patch
    "s2_original_name": str,       # original Sentinel-2 filename
    "s1_vh_name":       str,       # original Sentinel-1 VH filename
    "s1_vv_name":       str,       # original Sentinel-1 VV filename
    "question_ids":     List[int], # list of question IDs for this image
    "questions":        List[str], # list of questions
    "answers":          List[str], # list of answers
    "types":            List[str], # question type labels (see below)
}

---

Question Types

TAMMI defines 5 question categories divided into 21 sub-types, automatically generated from BDTopo, TRI, BU20, CLC, and EMA geographic databases:

1. Presence Questions

Type	Description	Example
presence	Presence of a BDTopo object class	"Is there a road in the image?"
pres_flooding	Presence of flooding risk (TRI)	"Is this area prone to flooding?"
pres_mountains	Presence of a mountain range (EMA)	"Are there any mountains in the image?"

2. Quantity Questions

Type	Description	Example
count	Number of objects of a BDTopo class	"How many buildings are in the image?"
density	Area of objects relative to patch area (BDTopo)	"What is the religious buildings density?"
area	Area of a specific object (BDTopo)	"What is the area of the lake?"
percent_clc	Percentage of patch area covered by a CLC class	"What percentage of the area is wetland?"

3. Location Questions

Type	Description	Example
smallest_largest	Absolute position of an object in a 3×3 grid	"Where is the largest vegetation area?"

4. Classification Questions

Type	Description	Example
clc	Dominant CORINE Land Cover class	"Which land cover category occupies the largest area?"
urban	Urban classification from BU20 (City Center, Suburb, Isolated City, Outside Urban Unit)	"What is the urban classification of the area?"
vegetation	Dominant/smallest vegetation type (BDTopo)	"Which vegetation type occupies the smallest area?"
water	Dominant/smallest water body type (BDTopo)	"What type of water body occupies the largest area?"
department	French department of the patch	"To which department does the area belong?"
region	French region of the patch	"To which region does the area belong?"
lvl_flooding	Highest flood risk level: low, medium, or high (TRI)	"What is the flood risk level?"
type_flooding	Type of flood risk: River overflows, Runoff, Sea Flooding, GroundWater overflows (TRI)	"What is the nature of the flood risk?"
name_mountains	Name of the mountain range (EMA)	"What is the name of the mountain range?"

5. Relational Analysis Questions

Type	Description	Example
distance	Distance in meters between two objects or a position and an object	"What is the distance between the museum and the religious place?"
less_more_equal	Comparison of object counts between two classes	"Are there more buildings than roads?"
spatial_relation	Relative position of one object with respect to another (octagon slice)	"What is the relative position of the monument with respect to the river?"
nearest	Position of the nearest object to a given coordinate in a 3×3 grid	"Where is the closest road to (142, 221)?"

---

Annotation Sources

Questions are automatically generated from the following official geographic databases:

• BDTopo (IGN) — vectorial description of French territory (buildings, water areas, roads, etc.)
• TRI (Géorisques) — flooding risk zones
• BU20 (INSEE) — urban unit classifications
• CLC 2018 (Copernicus) — CORINE Land Cover, Level-3 (44 classes)
• EMA (European Environment Agency) — European Mountain Areas

---

S1/S2 Mosaics

The full Sentinel-1 and Sentinel-2 department-level mosaics are available in the companion repository:

Himech/TAMMI_S1S2

To retrieve the S1/S2 patch for a given sample, crop the mosaic for the corresponding department using path_s1 / path_s2 as the relative crop coordinates, and s1_vh_name / s1_vv_name / s2_original_name to identify the mosaic file.

---

Usage

from datasets import load_dataset

ds = load_dataset("Himech/TAMMI75")

# Access a sample
sample = ds["train"][0]
image  = sample["bdo"]          # PIL Image (VHR orthophoto)
qs     = sample["questions"]    # list of questions
ans    = sample["answers"]      # list of answers
types  = sample["types"]        # list of question types

# Iterate over Q&A pairs for one image
for q, a, t in zip(qs, ans, types):
    print(f"[{t}] Q: {q} -> A: {a}")

---

Baseline Results

The paper proposes MM-RSVQA, a VisualBERT-based model combining VHR, MS, and SAR modalities:

Model	Overall Accuracy	Average Accuracy
VHR only	52.22	61.94
MS RGB + VHR	54.41	64.72
MS + VHR	54.49	64.87
SAR + VHR	54.29	64.40
MS + SAR	54.70	65.15
MM-RSVQA (all 3)	55.11	65.56

---

License

• BDOrtho: IGN Open License
• Sentinel-1/2: Copernicus Open Access

---

Citation

@inproceedings{boussaid2025tammi,
  title     = {Visual Question Answering on Multiple Remote Sensing Image Modalities},
  author    = {Boussaid, Hichem and Tosato, Lucrezia and Weissgerber, Flora and Kurtz, Camille and Wendling, Laurent and Lobry, Sylvain},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year      = {2025}
}

This work is supported by ANR under the ANR-21-CE23-0011 project. Experiments were performed using HPC/AI resources provided by GENCI-IDRIS (Grant 2023-AD011012735R2).