Zero Shot Classification for Change Detection in Satellite Imagery

This research investigates the zero-shot classification using the Comparative Language-Image Pre-Training (CLIP) model for change detection in satellite imagery. Since traditional supervised learning methods require extensive labeled datasets for all classes of objects in satellite images, which are often scarce or unavailable, zero-shot learning offers a promising alternative. Through detailed analysis of three different satellite image datasets (LEVIR-CD, DSIFN, and S2 Looking), the study evaluates the ability of the zero-shot classification model to identify changes without prior exposure to specific target classes and compares it with traditional tree-based supervised learning algorithms. These findings highlight the potential of zero-shot learning as a powerful tool for monitoring, managing, and responding to global changes, underscoring its importance for remote sensing and Earth observation applications where fast, accurate change detection is critical. The research is accompanied by the data and reproducible code at Github repository
¹
.