Semantic Retrieval and Multi-Temporal Change Analysis of Satellite lmagery.

Official description

Description 2.1. Background. Earth-observation archives are expanding rapidly and increasingly contain multi-temporal, multi-spectral and multi-sensor imagery. Conventional catalogues are effective for searching by metadata such as coordinates, acquisition date, platform and product type, but analysts may still need to know where and when to look before they can examine the imagery itself. Recent advances in multimodal and remote-sensing foundation models have improved semantic representation of Earth-observation data, creating the possibility of searching imagery by meaning as well as metadata. Translating those advances into a reliable operational system remains challenging. particularly when the system must work on-premises, ingest new acquisitions incrementally, preserve geospatial provenance, and suppress false change caused by season, atmosphere, viewing geometry, registration error or sensor differences. 2.2. Detailed Description. Teams are required to build a system that makes a satellite-imagery archive queryable by semantic content and by change over time, while retaining conventional spatial, temporal and sensor filters. The solution should support analyst discovery rather than require the analyst to identify every location of interest in advance. Six core capabilities are required. 2.2.1. Semantic and Multimodal Retrieval. Support free-text search over imagery tiles using natural-language queries, together with image-to-image search for visually and semantically similar locations. Results should be rank ordered and may be refined using area-of-interest, date-range, sensor or other metadata filters. Example queries include "newly built structures near a river" and "large vehicle concentrations on open ground". 2.2.2. Multi-Temporal Change Analysis. For a specified area and time window, identify meaningful changes such as appearance, disappearance, expansion or contraction of features; classify supported change types such as construction, clearance, water-extent variation or road development; and estimate the earliest available observation at which the change is supported by usable imagery. 2.2.3. False-Alarm Suppression and Quality Handling. Seasonal variation, illumination and view-angle differences, cloud, haze, snow, shadows, radiometric inconsistency and imperfect co-registration must be treated as confounding factors rather than automatically reported as change. The system should use quality masks, normalization, confidence estimates or equivalent mechanisms and should favour analytically useful precision over indiscriminate change recall. 2.2.4. Discovery and Clustering. Support unsupervised or embedding-based grouping of similar sites across a wider area so that an analyst who identifies one location of interest can discover other locations with comparable visual or semantic characteristics without manually constructing a new query for each site. 2.2.5. Analyst Workflow and Provenance. Provide a ranked review queue with before-and-after evidence, location, acquisition time, sensor or source information, confidence and relevant processing history. Analysts should be able to confirm or reject candidates, preserve those decisions in the audit trail, and use feedback for subsequent reranking or refinement where the chosen approach supports it. Exported results must retain source-scene and processing provenance. 2.2.6. Scale, Incremental Ingestion and Sovereignty. Support efficient vector or equivalent indexing, incremental addition of newly acquired imagery without a complete index rebuild, and complete on-premises operation without cloud services or external APIs during evaluation. Georeferencing and acquisition metadata must be preserved, and the solution should ingest organiser-defined common geospatial formats such as GeoTIFF or Cloud Optimized GeoTIFF (COG). 2.2.7. Constraints. The complete demonstration must run with network access. disabled after all approved models, libraries and datasets have been staged. locally. Pretrained public models may be used provided that their origin and licence are declared and the required weights are packaged for offline use. The evaluation will use publicly available or organiser-generated imagery only; no classified, operational or service-generated imagery will be included. 2.3 Expected Solution. A working system will be demonstrated over an organiser-defined area of interest and time span using public imagery. Retrieval will be evaluated against held-out semantic queries and relevance judgements, while change analysis will be evaluated against a held-out set of labelled change and no-change cases that participating teams have not seen. Teams must submit source code, an architecture note, the index-build and incremental-ingestion procedure, model and dataset provenance, and a reproducible evaluation report stating the indexed area, number of scenes or tiles, build time, storage footprint, query latency and hardware used. .