EcoMapper: Generative Modeling for Climate-Aware Satellite Imagery
Authors: Muhammed Goktepe, Amir Hossein Shamseddin, Erencan Uysal, Javier Muinelo Monteagudo, Lukas Drees, Aysim Toker, Senthold Asseng, Malte Von Bloh
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce a novel dataset of 2.9 million Sentinel-2 images spanning 15 land cover types with corresponding climate records, forming the foundation for two satellite image generation approaches using fine-tuned Stable Diffusion 3 models. ... We evaluate two generative models for their ability to integrate climate metadata into satellite image synthesis: Stable Diffusion 3 (SD3)... Diffusion Sat... We compare multiple configurations of Stable Diffusion 3 and Diffusion Sat, with and without fine-tuning, to assess their capacity for climate-aware satellite image synthesis. The qualitative results in Fig. 3 demonstrate the capability of our models in generating realistic satellite images conditioned on geographic and land type metadata. Table 1. Quantitative comparison of text-to-image generation models. |
| Researcher Affiliation | Academia | 1Technical University of Munich, School of Computation, Information and Technology, Germany 2University of Zurich, Department of Mathematical Modeling and Machine Learning, Eco Vision Lab, Switzerland 3Technical University of Munich, School of Computation, Information and Technology, Dynamic Vision and Learning Group, Germany 4Technical University of Munich, School of Life Sciences, Department of Life Science Engineering, HEF World Agricultural Systems Center, Chair of Digital Agriculture, Germany. |
| Pseudocode | No | The paper describes the methodology in narrative text and mathematical equations, but it does not include explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available under: https://github.com/maltevb/ecomapper. |
| Open Datasets | Yes | We introduce a novel dataset of 2.9 million Sentinel-2 images ... The dataset is available under: doi:10.14459/ 2025mp1767651 |
| Dataset Splits | Yes | The training set contains 98,930 locations, each spanning 24 months of data, while the test set includes 5,494 locations, each covering 96 months. Tab. 6 has details about the structure. The test set consists of 5,500 unique geographic locations, each monitored monthly over a 96-month period from 2017 to 2024. |
| Hardware Specification | No | The paper mentions computational cost and resource limitations but does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for the experiments. |
| Software Dependencies | No | The paper mentions several models and platforms used (Stable Diffusion 3, Diffusion Sat, Control Net, CLIP, T5, Google Earth Engine, Sentinel Hub, NASA Power) but does not provide specific version numbers for software libraries or dependencies like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | For text-to-image generation... both models were trained at 512 512 resolution... SD3... was tested in a fine-tuned experiment at 1024 1024 resolution. ... We fine-tuned this modified all layers of the model i.e. 900 million parameters on our training set... for 2 epochs. During fine-tuning, the encoder, decoder, and the CLIP text encoder were kept fixed... For the SD3-FT model, we applied Lo RA... and the model was trained for 2 epochs on our training set. For SD3-FT-HR... we fine-tuned 70 % of the Transformer blocks... reshaped the training images from 512 to 1024 resolution... trained for 2 epochs on a dataset of approximately 2 million images. |