Abstract
Traditional Post Classification Comparison (PCC) methods in remote sensing often fail to account for the spatial and temporal dependencies inherent in multitemporal imagery. This paper presents a novel object-oriented change detection approach that leverages spatiotemporal relationships to enhance classification accuracy. By employing Markov Random Fields (MRF) for spatial context and class transition probabilities for temporal logic, the proposed method refines initial Maximum Likelihood (ML) classifications through an Iterated Conditional Mode (ICM) procedure. Experimental results on high-resolution QuickBird imagery demonstrate a significant improvement in accuracy (93%) and a dramatic reduction in false alarms compared to standard PCC techniques.
Introduction
Change detection is a critical task in remote sensing, involving the identification of surface variations over time. The most common approach, Post Classification Comparison (PCC), involves classifying two images independently and comparing the results. However, PCC is prone to errors because it ignores the spatiotemporal context: land-cover transitions follow logical patterns (temporal dependence), and neighboring pixels often belong to the same class (spatial dependence). In high-resolution imagery, these omissions lead to significant "salt-and-pepper" noise and misclassifications between spectrally similar objects, such as dark rooftops and water bodies.
Methodology
The proposed framework addresses these limitations through an object-oriented pipeline designed to incorporate contextual information.
The process begins with Image Segmentation using a marker-based watershed algorithm to group pixels into homogeneous "objects." This shift from pixels to objects reduces noise and provides a more robust unit for analysis. Following segmentation, an Initial Classification is performed using a Maximum Likelihood (ML) method.
The core of the innovation lies in the Spatiotemporal Refinement stage. Spatial context is modeled using a Markov Random Field (MRF), which encourages neighboring objects to share the same class label. Temporal dependence is captured via a Transition Probability Matrix, which quantifies the likelihood of specific land-cover changes (e.g., farmland to building) over time. Finally, the Iterated Conditional Mode (ICM) algorithm is used to optimize the classification maps for both dates simultaneously, ensuring they are consistent with both the spectral data and the spatiotemporal rules.
Experimental Results
The method was tested on a multitemporal dataset of Wuhan, China, captured by the QuickBird satellite in 2002 and 2005. The study focused on five classes: lake, building, road, farmland, and bare land.
The proposed method achieved an overall accuracy of 0.93, a substantial improvement over the standard PCC method, which yielded accuracies of 0.71 and 0.66 for the respective years. The refinement process was particularly effective at suppressing false alarms, which dropped from 142 to just 13. By considering the temporal likelihood of transitions, the model successfully distinguished between spectrally similar classes, such as bare land and buildings, which often confuse independent classifiers.
Conclusion
By integrating spatial and temporal context into an object-oriented framework, this research significantly advances the state of change detection in high-resolution remote sensing. While the current model relies on supervised training and spectral data, the authors suggest that future work could incorporate texture and shape features to further enhance object recognition and explore unsupervised optimization techniques.
Summary updated