Recent advances in image-based satellite 3D reconstruction have progressed along two complementary directions. On one hand, multi-date approaches using NeRF or Gaussian-splatting jointly model appearance and geometry across many acquisitions, achieving accurate reconstructions on opportunistic imagery with numerous observations. On the other hand, classical stereoscopic reconstruction pipelines deliver robust and scalable results for simultaneous or quasi-simultaneous image pairs.

However, when the two images are captured months apart, strong seasonal, illumination, and shadow changes violate standard stereoscopic assumptions, causing existing pipelines to fail. This work presents the first Diachronic Stereo Matching method for satellite imagery, enabling reliable 3D reconstruction from temporally distant pairs.

Two advances make this possible: (1) fine-tuning a state-of-the-art deep stereo network that leverages monocular depth priors, and (2) exposing it to a dataset specifically curated to include a diverse set of diachronic image pairs. In particular, we start from a pretrained MonSter model, trained initially on a mix of synthetic and real datasets such as SceneFlow and KITTI, and fine-tune it on a set of stereo pairs derived from the DFC2019 remote sensing challenge. This dataset contains both synchronic and diachronic pairs under diverse seasonal and illumination conditions.

Experiments on multi-date WorldView-3 imagery demonstrate that our approach consistently surpasses classical pipelines and unadapted deep stereo models on both synchronic and diachronic settings. Fine-tuning on temporally diverse images, together with monocular priors, proves essential for enabling 3D reconstruction from previously incompatible acquisition dates.

We rely on MonSter, a state-of-the-art deep stereo model that incorporates monocular priors to improve performance in ill-posed scenarios. To adapt it to satellite diachronic pairs, we fine-tune it on a curated dataset of multi-date stereo pairs from the DFC2019 challenge, which includes a variety of seasonal and illumination changes.

Fine-tuning MonSter for diachronic stereo involves two key steps:

1) Curate a diachronic stereo dataset. We start from DFC2019 Track 3 with LiDAR-DSM ground truth over 110 AOIs (54 JAX, 56 OMA). Pairs are labeled by time gap and SIFT matches:

• Diachronic: >30 days apart and <40 SIFT matches.
• Synchronic: ≤30 days apart and ≥40 SIFT matches.

Comparison of SIFT and DISK+LightGlue matches for the diachronic pair OMA 331 017 – OMA 331 036. SIFT is not robust for finding matches in diachronic pairs.

To emphasize seasonal change, we sample >=30 diachronic pairs per OMA AOI and >=3 per JAX AOI, plus 5 random synchronic pairs per site. For each pair, we rectify the images and compute a ground-truth disparity using the reference DSM and RPC camera models. This yields 2,246 stereo pairs for the diachronic + synchronic dataset. Additionally, we also release a synchronic-only dataset with 1,567 pairs for ablations.

We release all data at huggingface.co/datasets/emasquil/diachronic-stereo, including diachronic+synchronic, synchronic-only, and the test sets used in the paper.

2) Rectify multi-date pairs for deep stereo. MonSter expects unipolar disparity (right image shifts to the left) that grows with altitude, but diachronic appearance changes make reliable matches that are needed for rectification scarce. We therefore use a dedicated rectification strategy (see Algorithm 1 below) that enforces these geometric constraints and outputs pairs compatible with the models.

We release the weights for the diachronic model at huggingface.co/emasquil/diachronic-stereo. The checkpoint should be loaded in a MonSter instance; see github.com/Junda24/MonSter.

We show qualitative results for disparities and DSMs on hard diachronic pairs. Classical pipelines and zero-shot deep models often fail under large appearance changes, producing noisy or incomplete geometry. Our method remains stable, yielding cleaner disparities and more accurate DSMs, which is reflected in consistently lower LiDAR MAE across test regions.

Figure 4. Qualitative disparity results on challenging diachronic pairs.

Figure 5. Top to bottom: DSMs from the classical s2p-hd pipeline, zero-shot MonSter (mix all), our method, and the ground-truth DSM. Missing values shown in black.