Introduction
Detecting changes in high-resolution remote sensing imagery is critical for urban planning, environmental monitoring, and disaster management. However, as image resolution increases, traditional methods struggle with "noise"—isolated pixels incorrectly identified as changes due to lighting variations or sensor artifacts. This paper, published in IEEE Access, proposes a hybrid framework that integrates pixel-level precision with object-level structural logic to enhance detection accuracy and reduce noise.
The Challenge of Noise in High-Resolution Data
Traditional pixel-level change detection methods often produce a "salt-and-pepper" effect, where minor spectral variations lead to numerous false positives. Conversely, object-level methods group pixels into meaningful entities (e.g., buildings, forests) but are highly dependent on the quality of initial segmentation and can be computationally intensive. The authors aim to bridge this gap by combining both approaches to capture the best of both worlds.
Methodology
The proposed workflow utilizes a dual-stream approach to leverage the strengths of both pixel and object-level analysis.
The process involves three primary stages:
- Pixel-Level Detection via Random Forest: The system extracts both spectral data and texture features (using a Gray-level Co-occurrence Matrix) to feed into a Random Forest classifier. This generates an initial, though potentially noisy, change map.
- Object-Level Segmentation via Improved U-Net: To define the "objects" in the scene, the team used an enhanced U-Net architecture. Improvements include Batch Normalization for stability, Dropout layers to prevent overfitting, and an increased number of convolution kernels to capture complex features.
- Adaptive Fusion and Refinement: This is the core innovation. The noisy pixel-level results are fused with the clean object boundaries from the U-Net. An adaptive threshold determines if an object has truly changed based on the density of "changed" pixels within its boundaries.
Datasets and Evaluation
The model was trained and tested on diverse high-resolution datasets, including the CCF Big Data Competition set (featuring images up to 8000x8000 pixels) and specific urban/agricultural test sets with resolutions ranging from 1.5m to 2.5m.
Key Results and Insights
The hybrid approach demonstrated significant improvements over traditional pixel-based methods like Principal Component Analysis and standard Difference Methods:
- Accuracy: The method achieved an Overall Accuracy (OA) of 92.72% on the first test set, compared to just 69.81% for standard difference methods.
- Precision: The False Alarm (FA) rate was reduced from 75.46% to 29.73%, indicating a much higher selectivity.
- Visual Quality: The resulting change maps were cleaner and more continuous, effectively eliminating the scattered false positives common in pixel-only approaches.
The inclusion of texture features (GLCM) proved vital for distinguishing between objects with similar spectral signatures but different physical structures, such as different types of vegetation or building materials.
Conclusion
By using semantic segmentation as a structural "mask" for pixel-level detections, the authors successfully mitigated the noise issues inherent in high-resolution imagery. While the current model identifies where changes occur, future work will focus on evolving the U-Net architecture to perform full "from-to" semantic change detection, providing a more descriptive narrative of land-use evolution.
Summary updated