Graph Neural Network-Based Method of Spatiotemporal Land Cover Mapping Using Satellite Imagery

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Machine Learning

Background:

• Accurate land cover classification is crucial for environmental monitoring and spatial modeling.
• Object-based methods have advanced satellite image classification, but spatiotemporal analysis remains challenging.
• Multispectral satellite imagery provides rich data for understanding land cover dynamics.

Purpose of the Study:

• To introduce a novel spatiotemporal method for object-based land cover classification using Graph Neural Networks (GNNs).
• To represent sequential satellite images as a directed graph, connecting segmented regions over time.
• To evaluate the method's performance against state-of-the-art models for intermonthly land cover mapping.

Main Methods:

• Utilized a modular pipeline with Convolutional Neural Networks (CNNs) for feature extraction (EfficientNetV2-S) and GNNs (GraphSAGE with LSTM aggregation) for node classification.
• Represented time-series satellite imagery as a directed graph where nodes are segmented land regions and edges represent temporal connections.
• Applied the method to Sentinel-2 L2A imagery for 4-year intermonthly land cover classification in Austria and Slovenia.

Main Results:

• The proposed GNN method outperformed the UNet model in land cover classification accuracy and F1-score for the Graz region (Level 2).
• Achieved superior performance over UNet in Level 1 classification (fewer classes) for both study regions.
• Demonstrated high classification accuracy for individual classes, reaching up to 99.17%.

Conclusions:

• The novel spatiotemporal GNN approach offers a significant advancement in object-based land cover classification from satellite imagery.
• The graph representation effectively captures land cover changes over time, leading to improved classification accuracy.
• This method provides a robust tool for generating detailed, intermonthly land cover maps for diverse geographical areas.