Automatic Planimetric Feature Extraction based on Model-Based Image Analysis, MBIA from simulated images
M.Sc. in Surveying Engineering, Remote Sensing, K.N.Toosi University of Technology
Tel. No.: +98 0451 7718120
Email: [email protected]
Obtaining information related to mapping is a common challenge for all world governments. For this purpose, feature extraction and the processing of the available remotely sensed data are major tools. In this paper an automatic model-based approach for planimetric feature extraction is presented.
The model-based approach is mapped to algorithms and software codes with the following modules and functions:
- Generate likelihood vectors (probability for radiometry given radiometric class) from multi-spectral remote sensing data and a radiometric model. Store the likelihood vectors in radiometric evidence maps.
- Generate geometric hypothesis given initial parameters and with geometric constraints. Include the signal mixture model as defined by sensors point-spread function. Store each iteration over the geometric parameter space in a probability for class vector hypothesis map.
- Evaluate each iteration of the hypothesis generation process calculating the degree of agreement or disagreement between the evidence map and the hypothesis map. This result is the average likelihood over all classes and all objects. When multiplied with a cost/benefit matrix a monetary utility function is produced.
- If the utility is maximum then stop and store this.
- If the utility is less than maximum, then change the model (parameter) and go to step 2 and generate new geometric hypothesis maps.
A typical feature of the method is the separation of the complex combined probabilities of multi-spectral data, object class, object geometry and sensing into a radiometric model and a geometric model. Bayesian inversion of the radiometric model leads to radiometric evidence maps. The geometric hypothesis maps are generated from parameterized object models. Geometric models are constrained by fixed objects.
The goal of model-based approach is to optimize a utility function, which can be described as cost- or risk-weighted likelihood for a collection of objects and their parameters. The utility function consists, for each iteration of geometric hypotheses generation, of the matrix over the crossing of all evidence vectors and all hypotheses vectors. This “confusion” matrix is multiplied with a benefit/cost matrix. Then study of the utility (parameter) function gives the overall quality in the minimum value of utility. Knowledge about spectral properties of objects is already handled through the Bayesian inversion of the probability for spectral value given class. Knowledge about the geometry of objects can be obtained directly from the data through likelihood-based segmentation and/or from an existing geographic object model. Knowledge about remote sensing, such as knowledge contained in sensor models and atmospheric transmission models, can be added in one conditional probability model.
This research has produced findings that some of them are as follows:
- Top-Down image analysis by hypotheses generation using a parameterized world model and by varying the model parameters.
- Edge detection and parameter estimation given a geometric model
- Per-object cost/benefit analysis
- Likelihood-based segmentation
The efficiency of the model-based approach is compared with maximum likelihood classification and fuzzy classification and it has promising results and because model-based theory is newly born, it is in its early stages and we think that in early future, model-based methods will play an important role in image processing and feature extraction.