InSAR+: Exploring the utility of complementary data sources for mapping conflict damage using InSAR coherence time series

Reliable and timely damage assessments are critical in humanitarian and conflict settings. The event-based analysis of very high-resolution (VHR) optical imagery has been the predominant remote-sensing based method to achieve this, but remains constrained by limited temporal revisit, cloud cover, cost, and restricted spatial scalability. Interferometric Synthetic Aperture Radar (InSAR) coherence derived from Sentinel-1 offers a complementary, medium-resolution approach by enabling frequent, weather-independent observations over large areas, making it particularly suitable for near-real-time and retrospective damage monitoring. However, the potential of InSAR coherence time series remains underexplored, particularly in how it can be complemented by other sensors (e.g., optical imagery) and how it is affected by different built-up environment characteristics.

 

This study investigates large-scale conflict-related damage mapping across Gaza during 2023–2024 using Sentinel-1 InSAR coherence time series. We also integrate multiple data sources, including Sentinel-2 optical imagery, gridded weather re-analysis data, and built-up environment characteristics. Moreover, we generate embeddings of the Sentinel imagery using geospatial foundation models which we use as additional model inputs. Damage reference data are derived from UNOSAT damage assessments, which report damage at irregular intervals (~2-3 months) based on visual assessments of VHR optical imagery. To exploit the higher temporal frequency of Sentinel-1 acquisitions while accounting for the coarser temporal resolution of the reference data, we adopt a weakly supervised multiple instance learning framework and compare the predictive performance of our model across various combinations of input modalities.

 

The analysis aims to quantify the relative importance of different input modalities for damage detection, assess the added value of self-supervised representation learning, and identify inherent limitations related to site-specific, sensor-specific and damage-specific factors in Gaza. We further evaluate the utility of interval-based learning approaches for conflict damage monitoring, where precise damage timing is often unavailable.

 

By combining dense SAR time series, multimodal data fusion, and interval-aware learning, this work contributes a novel methodological perspective on large-scale damage assessment. The findings inform both the potential and limitations of InSAR-based damage mapping in humanitarian contexts, supporting future operational monitoring and post-event re-analysis workflows.