General-domain text-guided image editors achieve strong photorealism but introduce artifacts, hallucinate objects, and break the orthographic constraints of remote sensing (RS) imagery. We trace this gap to two high-level causes: (i) limited RS world knowledge in pre-trained models, and (ii) conditioning schemes that misalign with the bi-temporal structure and spatial priors of Earth observation data. We present RSEdit, a unified framework that adapts pretrained text-to-image diffusion models—both U-Net and DiT—into instruction-following RS editors via channel concatenation and in-context token concatenation. Trained on over 60,000 semantically rich bi-temporal remote sensing image pairs, RSEdit learns precise, physically coherent edits while preserving geospatial content. Experiments show clear gains over general and commercial baselines, demonstrating strong generalizability across diverse scenarios including disaster impacts, urban growth, and seasonal shifts, positioning RSEdit as a robust data engine for downstream analysis. We will release code, pretrained models, evaluation protocols, training logs, and generated results for full reproducibility.
@misc{rsedit2026,
author = {Zhenyuan Chen, Zechuan Zhang, Feng Zhang},
title = {RSEdit: Text-Guided Image Editing for Remote Sensing},
howpublished = {\url{https://github.com/Bili-Sakura/RSEdit-Preview}},
url={https://arxiv.org/abs/2603.13708},
year = {2026}
}
Overview of the RSEdit framework. We propose a universal adaptation strategy that aligns the conditioning mechanism with the architecture’s inductive bias. For U-Net backbones (left), we use channel concatenation to leverage convolutional priors. For DiT backbones (right), we use token concatenation to exploit the in-context learning capabilities of transformers.
Proposed change-centric evaluation metric using building damage assessment. We leverage a pre-trained ChangeStar model to produce semantic masks for the edited image and compare them to ground-truth damage masks to quantify editing accuracy.
Table 1: Change detection evaluation results. $F1_{\text{dam}}$ denotes the harmonic mean of per-class F1 scores following xView2 protocol. Bold indicates best results, underline indicates second best.
| Method | $F1_{\text{dam}} \uparrow$ | VIE Scores $\uparrow$ | ||
|---|---|---|---|---|
| SC | PQ | $\text{VIE}_{\text{overall}}$ | ||
| SD 1.5$^\dagger$ | 6.87 | 3.332 | 2.457 | 2.492 |
| SD 2.1$^\dagger$ | 6.05 | 2.958 | 2.042 | 2.120 |
| Text2Earth$^\dagger$ | 5.06 | 3.970 | 3.494 | 3.229 |
| InstructPix2Pix | 8.37 | 4.462 | 3.195 | 3.153 |
| MagicBrush | 0.96 | 2.205 | 2.006 | 1.638 |
| UltraEdit | 1.16 | 4.250 | 2.098 | 2.531 |
| Flux.1-Kontext | 5.41 | 5.065 | 4.047 | 3.693 |
| RSEditU-Net | 34.11 | 5.793 | 3.657 | 4.129 |
| RSEditDiT | 25.94 | 5.759 | 4.024 | 4.210 |
Out-of-domain generalization on LEVIR-CC and SECOND-CC test sets. Our model trained on RSCC generalizes well to unseen benchmarks without fine-tuning.
Table 3: Out-of-domain generalization on LEVIR-CC [Liu et al. 2022] and SECOND-CC [Karaca et al. 2025] test sets. Our model trained on RSCC generalizes well to unseen benchmarks without fine-tuning.
| Method | SC $\uparrow$ | PQ $\uparrow$ | VIE $\uparrow$ |
|---|---|---|---|
| LEVIR-CC | |||
| InstructPix2Pix | 4.98 | 2.33 | 2.71 |
| UltraEdit | 2.45 | 1.28 | 1.18 |
| RSEditU-Net | 3.88 | 3.97 | 2.74 |
| RSEditDiT | 3.85 | 4.41 | 2.82 |
| SECOND-CC | |||
| InstructPix2Pix | 3.94 | 1.87 | 1.96 |
| UltraEdit | 3.69 | 1.17 | 1.57 |
| RSEditU-Net | 4.28 | 2.67 | 2.61 |
| RSEditDiT | 4.60 | 3.00 | 2.91 |
High-resolution results for disaster scenario editing. Each figure displays RSEdit's editing capabilities across different disaster scenarios.
Table 3: Full text instructions for the samples used in qualitative analysis.
| Scenario | Instruction |
|---|---|
| Guatemala Volcano | The volcanic eruption transformed the lush golf course into a desolate ash-covered landscape, with once-pristine fairways blanketed in gray debris and expanded water bodies inundating low-lying areas. Five buildings exhibit minor damage, showing partial burns, cracked roofs, and scattered volcanic material, while six structures suffer major collapse, their walls and roofs shattered by encroaching lava flows or mudslides. Two properties are entirely obliterated, reduced to scorched foundations or submerged under torrents of water and sediment, marking a catastrophic shift from pre-event tranquility to post-disaster devastation. |
| Joplin Tornado | The previously intact residential area has been transformed into a scene of widespread devastation, with 43 buildings entirely obliterated—their foundations scorched, collapsed, or submerged in mud—while 8 structures suffered partial roof or wall collapses and debris encroachment, and 3 exhibited cracked facades or localized burn marks, leaving no properties untouched amidst the catastrophic wreckage. |
| Portugal Wildfire | Wildfire impacts revealed minor damage to three buildings, showing partial burn marks, scattered debris, and disrupted roofing, while one structure remained undisturbed. Surrounding vegetation exhibited widespread charring, with ash-covered patches and reduced greenery density. Roads displayed residual smoke haze and localized erosion from firefighting water runoff, altering terrain contours near affected sites. |
| Hurricane Florence | Severe flooding engulfed the area, submerging all six buildings up to their rooftops, causing partial wall collapses and significant structural weakening. Vegetation along the shoreline was stripped away by rushing waters, exposing bare earth and debris. Roads near the settlement became impassable due to mudslides and erosion, isolating the community. No intact structures remained visible, with every building classified as majorly damaged (Level 2) under disaster protocols. |
| Mexico Earthquake | Post-earthquake assessment shows no structural damage detected in 89 evaluated buildings, with all structures maintaining intact roofs, walls, and foundations. Nearby infrastructure, including roads and vegetation, displays no evidence of collapse, flooding, or volcanic activity. The area visually aligns with pre-disaster conditions, resulting in Disaster Level 0 (No Damage) classification for all inspected properties. |
| Palu Tsunami | A devastating tsunami has caused catastrophic destruction across the coastal settlement, transforming the landscape from a densely populated area into a scene of widespread ruin. Pre-event imagery revealed 82 intact structures nestled along the shoreline, while post-event analysis reveals 66 buildings obliterated (Disaster Level 3)—scorched, submerged, or entirely erased—leaving only skeletal remains or vacant lots. Seven properties suffered major damage (Level 2), with collapsed roofs, encroaching mudflows, or partial wall failures, concentrated near the waterfront. Nine buildings escaped unscathed (Level 0), situated on elevated terrain or shielded from surge impacts. Coastal erosion reshaped the shoreline, displacing boats and scattering debris, as turbid waters inundated low-lying zones, marking the tsunami's merciless advance. |