Abstract:We consider a stationary and isotropic spatial point process, whose a realisation is observed within a large window. In order to predict its local intensity, we propose to define the first- and second-order characteristics of a random field, defined as the regularized counting process, from the ones of the point process and to interpolate the intensity by using a revisited kriging of the regularized process.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to predict the local intensity of partially observed data, especially in areas outside the observation window. Specifically, the author focuses on how to use the re - defined Kriging method to predict the local intensity of a stationary and isotropic spatial point process in unobserved areas. This problem is particularly common in plant ecological sampling because the sampling areas are often discontinuous.
### Background and Motivation of the Paper
When estimating the intensity of a point process, it is usually necessary to observe the complete point pattern within a given window and understand its local variation on a given grid. Existing methods include kernel smoothing, parametric methods, etc., but these methods have difficulties in choosing the bandwidth or grid size. In addition, these methods mainly focus on the intensity change within the observation window and have limited ability to predict the intensity outside the observation window.
### Paper Method
The author proposes a new Kriging method. By regularizing the counting process of the point process into a random field, and then using the first - and second - order characteristics of this random field to interpolate the local intensity. The specific steps are as follows:
1. **Regularize the Counting Process**: Define a random field \(Z(x)\) as the number of points of the point process in the grid cell \(B\) centered at \(x\), that is, \(Z(x)=\Phi(x\oplus B)\).
2. **Link the Characteristics of the Point Process and the Random Field**: Describe the mean and variogram of the random field through the intensity \(\lambda\) of the point process and the pair - correlation function \(g(r)\).
3. **Kriging Interpolation**: Use the ordinary Kriging method to predict the local intensity \(\hat{\lambda}(x_0|\Phi_{S_{\text{obs}}})\) at the unobserved position \(x_0\) according to the observed data \(\Phi(B_i)\).
### Paper Contributions
1. **Expand the Scope of Application**: This method is not only applicable to Cox processes, but also to a broader class of point processes.
2. **Optimize the Scale**: Determine the optimal grid size by minimizing the Integrated Mean Squared Error (IMSE).
3. **Statistical Properties**: When the pair - correlation function is known, this method has good statistical properties such as unbiasedness and low variance. Even when the pair - correlation function needs to be estimated, the performance of this method is relatively good.
4. **Practical Application**: Through the actual case of the Montagu's harrier nest sites, the effectiveness of this method in predicting the local intensity of unobserved areas is demonstrated.
### Experimental Results
The author verifies the effectiveness of the method through simulation experiments. The results show that:
- When the pair - correlation function is known, the prediction results are unbiased and have low variance.
- When the pair - correlation function needs to be estimated, the accuracy of the prediction results decreases slightly, but it can still effectively predict the local intensity, especially when the observation points are close to the boundary of the unobserved area.
### Conclusion
The method proposed in this paper performs well in predicting the local intensity of partially observed data, especially for highly clustered small - scale point processes. This method is not only applicable to Cox processes, but also to a broader class of point processes and has broad application prospects.