Abstract
Geometric accuracy of the remote sensing rectified image is usually evaluated by the root-mean-square errors (RMSEs) of the ground control points (GCPs) and check points (CPs). These discrete geometric accuracy index data represent only a local quality of the image with statistical methods. In addition, the traditional methods only evaluate the difference between the rectified image and reference image, ignoring the degree of the original image distortion. A new method of geometric quality evaluation of remote sensing image based on the information entropy is proposed in this paper. The information entropy, the amount of information and the uncertainty interval of the image before and after rectification are deduced according to the information theory. Four kinds of rectification models and seven situations of GCP distribution are applied on the remotely sensed imagery in the experiments. The effective factors of the geometrical accuracy are analysed and the geometric qualities of the image are evaluated in various situations. Results show that the proposed method can be used to evaluate the rectification model, the distribution model of GCPs and the uncertainty of the remotely sensed imagery, and is an effective and objective assessment method.
Introduction
Geometric rectification is a critical process in remote sensing that aligns satellite imagery with real-world geographic coordinates. Traditionally, the quality of this rectification is assessed using the Root-Mean-Square Error (RMSE) at specific Ground Control Points (GCPs) and Check Points (CPs). However, RMSE-based evaluation has notable limitations: it provides only a local statistical measure, often relies on subjective confidence levels, and fails to account for the initial distortion state of the raw image. This research proposes a novel framework based on information entropy to provide a more objective and global assessment of geometric quality.
Methodology
The proposed method utilizes Shannon Entropy to quantify spatial uncertainty. The framework calculates several key metrics:
- Priori Entropy ($H(X)$): The initial uncertainty of image coordinates before rectification.
- Posteriori Entropy ($H(X|X_n)$): The remaining uncertainty after applying a rectification model.
- Information Gain ($I$): The difference between the two ($I = H(X) - H(X|X_n)$), representing the reduction in spatial disorder achieved through the rectification process.
Additionally, the authors derived an Uncertainty Interval ($\Delta$). By applying properties of information theory to a normal distribution of errors, they established a fixed coefficient ($k = 2.066$), resulting in an objective error range of $\pm 2.066\sigma$. This removes the need for researchers to subjectively choose confidence thresholds (e.g., 95% or 99%).
Experimental Setup
The study utilized Landsat-5 TM imagery of the Aksu district in Xinjiang, China. This region features challenging terrain with elevations ranging from 600m to 4000m. To test the robustness of various models, the researchers designed seven different spatial distributions of GCPs (labeled a-g), ranging from well-distributed patterns to localized clusters.
Four rectification models were compared: a Rigorous Physical Model, a second-degree Polynomial Model, a Rational Function Model (RFM) with Ridge Estimation, and a robust RFM using $l_1$-norm Regularized Least Squares (LS1).
Results and Discussion
The entropy-based approach revealed critical insights into model stability. A significant finding was that the Information Gain ($I$) can be negative. In scenarios with poor GCP distribution (such as scenario 'c'), unstable models like RFM with ridge estimation actually increased the spatial uncertainty of the image, making it more disordered than the raw original.
Quantitatively, the RFM (LS1) model demonstrated superior performance across nearly all scenarios. In a well-distributed configuration (scenario 'f'), the uncertainty interval in the X-direction was reduced from 375.42 meters in the raw image to just 19.02 meters. Overall, the models were ranked by their ability to gain information as follows: RFM (LS1) > Rigorous Physical Model > Polynomial > RFM (Ridge Estimation).
The results also quantitatively confirmed that "interpolation" (where GCPs surround the study area) is far superior to "extrapolation" (where GCPs are clustered on one side). For instance, in scenario 'c', the RFM (ridge estimation) produced an extreme X-direction uncertainty of 2437.24 meters.
Conclusion
The information entropy method offers a robust and objective alternative to traditional RMSE for evaluating the geometric quality of remote sensing images. By treating geometric accuracy as a measure of uncertainty reduction, it provides a holistic view of model performance and spatial integrity. Future work will focus on applying this method to high-resolution imagery and integrating it into automated satellite data processing pipelines.
Summary updated