Land-Cover Change Detection

Vision AI Model

ONNX

Fully Convolutional Siamese Network

This application implements a Fully Convolutional Siamese Network (FCSN) for land-cover change detection. The model uses two parallel, weight-sharing convolutional encoders that independently extract hierarchical spatial features from a pair of multi-temporal satellite or aerial images. The resulting feature maps are combined through an absolute-difference operation, generating a discriminative representation of pixel-level change. A shared fully convolutional decoder processes these difference maps to reconstruct a dense change mask, enabling the model to identify meaningful land-cover transitions—such as natural disasters, deforestation, or urban expansion—while filtering out irrelevant variations. Edge-optimized quantization reduces memory footprint and latency to support real-time inference on SECO hardware, and post-processing modules refine the predicted segmentation masks and change heatmaps for downstream geospatial analysis.