Abstract
Image classification has advanced significantly over the last half-century, yet a gap remains between theoretical research and practical application. This paper explores the integration of artificial intelligence techniques, specifically Bayesian Networks, into the remote sensing domain to address uncertainty in complex environments. We apply Tree Augmented Naive Bayesian Networks (TAN) to the texture classification of high-resolution satellite imagery, specifically supporting China’s First National Geographical Information Census. By constructing network topologies based on training accuracy, we demonstrate that TAN outperforms traditional Naive Bayesian and Maximum Likelihood methods. Our results indicate that this approach improves image interpretation efficiency and reduces the workload for field verification.
Introduction: The Challenge of National Census Mapping
In 2013, the Chinese government initiated its first national geographical information census, a massive project relying on high-resolution remote sensing imagery to map land cover and land use. While high-resolution data offers unprecedented detail, it introduces a significant bottleneck for human interpreters. While humans are adept at defining the boundaries of land areas (segmentation), accurately classifying these areas—such as distinguishing between coniferous, broad-leaved, and mixed forests—remains difficult and often requires costly field verification. This research seeks to bridge this gap by using probabilistic AI to assist office-based interpreters.
Methodology: Tree Augmented Naive Bayesian Networks (TAN)
The core innovation of this study is the application of the Tree Augmented Naive Bayesian Network (TAN). Traditional Naive Bayesian Classifiers (NBC) assume that all image features, such as color and texture, are independent. However, in remote sensing, features like "roughness" and "regularity" are often highly correlated.
TAN improves upon NBC by allowing for a more complex topology where the class (e.g., "Forest") acts as a parent node, but dependencies are also modeled between child nodes (the features). We extracted two types of features for the network:
- Statistical Texture Features: Derived from Gray-Level Co-occurrence Matrices (GLCM), including entropy and angular second moment.
- Structural Texture Features: Captured using Symlets Wavelet transforms to identify underlying landscape patterns.
Experimental Setup and Datasets
The study utilized high-resolution colorized remote sensing images from the 2013 census in Beijing, with resolutions ranging from 0.19 meters in plains to 0.48 meters in mountainous areas. The experiment focused on the three most frequently confused classes: broad-leaved forest, coniferous forest, and mixed broadleaf-conifer forest.
The workflow employs a human-machine hybrid approach: a human operator manually outlines "irregular polygons" (objects), and the TAN model then calculates the most probable classification for that object based on its extracted features.
Key Results and Insights
The performance of TAN was compared against the Naive Bayesian Classifier (NBC) and the Maximum Likelihood Classification (MLC) method across various training sample sizes.
- Superior Accuracy: TAN consistently outperformed both NBC and MLC. With 30 training samples, TAN reached a peak accuracy of 91.4%, compared to 87.1% for NBC and 85.7% for MLC.
- Efficiency with Small Data: Even with only 15 training samples, TAN reached 88.24% accuracy, surpassing the performance of other methods using twice the data.
- Feature Correlation: The results confirm that capturing dependencies between remote sensing features significantly improves classification nuances in high-resolution textures.
Conclusion and Future Work
While the TAN model is more computationally intensive than simpler classifiers due to the need to construct complex network topologies, it offers a highly effective tool for assisting office-based image interpretation. By reducing the need for field verification, it significantly improves work efficiency for large-scale projects like the national census. Future research will focus on expanding the model's structure to handle the full range of seventy-seven land-use classes identified in the national census and optimizing computational performance for broader applications.
Summary updated