EcoMapper

April 27, 2026 · View on GitHub

EcoMapper is a generative modeling framework designed to produce realistic, climate-aware satellite imagery using advanced diffusion models. This project combines remote sensing with climate data, leveraging a globally distributed dataset of 2.9 million Sentinel-2 images across 15 land cover types and 6 years of monthly data.

Paper & Dataset

Our ICML 2025 paper is available on OpenReview:

📄 Paper: https://openreview.net/forum?id=YUtJsxQjv3

The paper introduces the EcoMapper framework, dataset creation pipeline, training methodology, and evaluation benchmarks for climate-aware satellite image generation.

The EcoMapper dataset is publicly available here:

🌍 Dataset Download: https://mediatum.ub.tum.de/node?id=1767651

The dataset contains:

2.9 million Sentinel-2 images
15 land cover classes
6 years of monthly observations
Global geographic coverage
Associated climate metadata for each image

Each satellite image is paired with a metadata .csv file that stores temporal, geographic, environmental, and land-cover information used for training and conditioning.

Metadata Format

Each row corresponds to a single satellite observation:

Column	Description
`location_id`	Unique identifier for a geographic location
`date`	Observation date in `YYYY-MM` format
`latitude`	Latitude coordinate
`longitude`	Longitude coordinate
`type`	Land cover class
`location_address`	Human-readable location
`cloud_coverage`	Cloud coverage value
`average_temp`	Average temperature
`average_prep`	Average precipitation
`average_rad`	Average solar radiation
`img_path`	Path to corresponding satellite image
`seed`	Dataset generation/split seed

Example Metadata Entry

location_id	date	latitude	longitude	type	cloud_coverage	average_temp	average_prep	average_rad	img_path
1	2017-02	62.99	43.42	Evergreen Needleleaf Forests	0.108	-10.52	1.84	1.61	`1_2017-02.png`

This metadata enables conditioning on:

Geographic location
Temporal changes
Climate variables
Land cover categories
Seasonal variations

Introduction

EcoMapper features two core capabilities:

Text-to-Image Generation: Uses Stable Diffusion 3 (SD3) to synthesize satellite images from climate-aware text prompts, incorporating location, time, land cover type, and environmental conditions such as temperature, precipitation, and solar radiation.
Multi-Conditional Image Generation: Integrates ControlNet to guide generation with both text and image inputs, preserving spatial structures while simulating environmental changes over time.

These models enable:

The creation of cloud-free synthetic observations for regions affected by persistent coverage
Simulation of future landscapes under different climate scenarios
Realistic input generation for forecasting tasks in agriculture, disaster response, and land cover monitoring

By fusing multimodal generative AI with Earth observation data, EcoMapper advances climate-aware modeling and opens up new avenues for scenario analysis, geospatial intelligence, and environmental visualization.

Framework Overview

Figure 1: Text- and image-conditioned satellite image generation using Stable Diffusion 3 and ControlNet.

Model Outputs Across Regions

Generation Comparison

Figure 2: Satellite images generated by fine-tuned models vs. ground truth across diverse locations and land cover types.

Climate Sensitivity in Generated Imagery

Climate Sensitivity

Figure 3: Effect of extreme climate inputs (e.g., dry boreal, humid tropical) on generated satellite imagery for different regions.

Installation

To get started, set up the environment and install the required dependencies by following the steps below.

First, clone the repository to your local machine and then create an environment to install requirements:

git clone https://github.com/maltevb/ecomapper.git
cd Ecomapper
conda create -n ecomapper python=3.10
conda activate ecomapper
pip install -r requirements.txt

Note: For diffusionsat related activity, visit https://github.com/samar-khanna/DiffusionSat and follow the corresponding instructions to create an appropriate environment.

Training

EcoMapper provides different versions for training models with or without LoRA, as well as for single image generation and ControlNet. For each training session, two configuration files are needed: one for model architecture, dataset and training parameters (config_train), and another for the masks used for prompting (config_caption).

Single Image Generation

You can train the model with or without LoRA by selecting the appropriate configuration files.

SD3 without LoRA

To train the model without LoRA, use the following command:

accelerate launch --mixed_precision=fp16 train.py \
    --config_train configs/train/train.yaml \
    --config_caption configs/caption/skip_date_location.json

SD3 with LoRA

To train the model with LoRA, use the following command:

accelerate launch --mixed_precision=fp16 train_lora.py \
    --config_train configs/train/train_lora.yaml \
    --config_caption configs/caption/skip_date_location.json

ControlNet Image Generation

For ControlNet image generation, the process is similar: you can train the model with or without LoRA by selecting the corresponding configuration files.

ControlNet SD3 without LoRA

To train ControlNet without LoRA, use this command:

accelerate launch --mixed_precision=fp16 train_controlnet.py \
    --config_train configs/train/train_controlnet.yaml \
    --config_caption configs/caption/skip_date_location.json

ControlNet SD3 with LoRA

To train ControlNet with LoRA, run this command:

accelerate launch --mixed_precision=fp16 train_controlnet_lora.py \
    --config_train configs/train/train_controlnet_lora.yaml \
    --config_caption configs/caption/skip_date_location.json

Inference

After training, use the validation script to generate results from the trained model. Ensure that you specify the correct model paths in the script.

Inference for ControlNet SD3 without LoRA

python validation/validate_controlnet.py

Inference for ControlNet SD3 with LoRA

python validation/validate_controlnet_lora.py

Computing the Metrics (after validation scripts)

Notes for diffsat metrics: To compute clip score, before running batch_compute_metrics.py, insert the clip caption from the json files of the SD3 images to the json file of the diffsat images.

SD3

python validation/batch_compute_metrics.py

ControlNet

python validation/batch_compute_metrics_controlnet.py

Citation

If you find our project helpful, please cite our paper:

@inproceedings{goktepe2025ecomapper,
  title     = {EcoMapper: Generative Modeling for Climate-Aware Satellite Imagery},
  author    = {G{\"o}ktepe, Muhammed and Shamseddin, Amir Hossein and Uysal, Erencan and Monteagudo, Javier and Drees, Lukas and Toker, Aysim and Asseng, Senthold and von Bloh, Malte},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year      = {2025},
  month     = {July},
}