Current techniques for interpreting imagery and populating GIS databases are inadequate. Manual interpretation is too slow, and current image processing techniques are generally not accurate enough. Recent improvements in image interpretation utilize inductive learning algorithms. These systems show promise because they can process imagery quickly. However, objects in images are very complex. It is difficult for inductive learners to identify complex features in an image with one model. The results are often cluttered. This paper presents a system that applies a hierarchy of inductive learning algorithms that assist an analyst in interactively removing classification errors through a "data-driven" process.
Business point-of-sale analyses, military intelligence operations, local government infrastructure planning, telecommunications, and many other industry applications all rely on geospatial data. As Geographic information system (GIS) applications proliferate, the laggard in the equation is the temporal currency and spatial resolution of the data itself. In fact, the process of updating GIS databases is estimated to be 60-80% of the cost of GIS (Oswald et al., 1998). The natural and predominate technique for populating GIS databases is by understanding and classifying remotely sensed images. Even though the bandwidth of digital imagery from space is steadily increasing, the opportunity to extract geospatial data from imagery is limited due to (a) the costs of manual digitizing methods, (b) the lack of trained image analysts, and (c) the inefficiency of automated feature extraction techniques. This article investigates the utility of the Feature Analyst (www.featureanalyst.com) to reduce the time and effort currently required for extracting geospatial features from digital imagery.
The feature extraction approach of the Feature Analyst does not follow the traditional automated-target-recognition /automated-feature-extraction (ATR/AFE) path. Rather it uses machine learning to bridge the gap between GIS and traditional image processing. The Feature Analyst works by having the user provide samples of extracted features from the image. The system then automatically develops a model that correlates known data (such as spectral signatures) with targeted outputs (i.e., the materials or objects of interest). The learned model subsequently classifies and extracts the remaining materials or objects.
A noticeable tendency of inductive learners (Quinlan, 1986; Mitchel, 1997) is to over predict the frequency of a feature in the image, sometimes called the "false positive effect" (Palhang, 1997). This means that the learner incorrectly includes in its classification superfluous features along with correctly identified features. It is a difficult task to identify all the nuances of a complex feature in a single classification. Feature Analyst addresses this problem by applying a hierarchy of inductive learners. This article examines the efficiency of the Feature Analyst on NIMA's Foundation Feature classes using various types of imagery.
Another option is to employ a computer specialist to write custom feature extracting software for a specific geographic area. These custom approaches may work well for a given geographic area, but often perform poorly if either the feature or image characteristics change. For instance, a road finding algorithm that works well with images taken in the Spring may need to be radically modified to find roads in the same area in the Winter.
The Feature Analyst extension for Esri's ArcView and ArcGIS software, developed by Visual Learning Systems (VLS), Inc. (www.featureanalyst.com), provides a paradigm shift to feature extraction since it: (a) utilizes spectral, spatial, temporal, and ancillary information to model the feature extraction process, (b) provides the ability to remove clutter, (c) incorporates advanced machine learning techniques, and (d) provides an exceedingly simple interface for feature extraction.
The primary point of our investigation is to determine the utility of the Feature Analyst on various NIMA Foundation Feature Data classes. Table 1 presents the features and associated image information. Opitz, 2002, used the same datasets when investigating the value of contextual classification. Ground truth was created by carefully hand digitizing each desired feature in the image sample.
Table 1. Feature and imagery used during the study.
| Feature | Image Type | Number of Bands | Pixel Width | Image Source |
| Airport | PAN/MSI | 3 | 1 m | Ikonos-2 |
| Buildings P | Pan | 1 | 1 m | Ikonos |
| Buildings M | MSI | 4 | 0.5 m | ADAR-5500 |
| Built-up Area P | Pan | 1 | 1 m | Ikonos |
| Built-up Area M | MSI | 7 | 10 m | Atlas |
| Coastline/Shoreline | MSI | 7 | 28.5 m | Landsat 5 |
| Ditch | Pan | 1 | 1 m | Ikonos |
| Port Facility | Pan/MSI | 3 | 1 m | Ikonos-2 |
| River Network P | Pan | 1 | 1 m | Ikonos |
| River Network M | Pan Sharpened | 3 | 1 m | Ikonos |
| Road Network | Pan Sharpened | 3 | 1 m | Ikonos |
The goal of the Feature Analyst's hierarchical feature extraction is to leverage a human's impressive vision ability to improve classification results by mitigating clutter (false positives), and retrieving missed objects. Hierarchical learning is necessary for learning complex themes from a diverse set of information. It effectively breaks down the learning task as well as naturally generates a large training set that addresses the problem breakdown. The overall process iteratively narrows the classification task into sub-problems that are more specific and well defined. Feature Analyst begins the hierarchical process the same as we approach any baseline inductive learning classification, i.e., select labeled examples for the feature being extracted, train the learner, and then classify every pixel in the image based on the learner's prediction. At this point if the users are not satisfied with the results, they can apply a hierarchy of learners to improve the classification. The classification is improved in sequential passes; each new pass is designed to remove one form of error from the results of the previous pass, while in the process, generating from the GUI an easily obtained set of training examples.
Tables 2 and 3 show the results from four passes of hierarchical learning. The second pass is clutter mitigation, the third pass includes missed-feature retrieval and the last pass is clutter mitigation. We found that, almost uniformly, whenever one brings in missed-features, at least one more pass of clutter mitigation is necessary. Note that hierarchical learning significantly increases the accuracy of the extracted features.
Table 2. Percent accuracy after multiple passes on the multispectral imagery.
| Feature | Pass 1 Accuracy | Pass 4 Accuracy |
| Building (Dark) | 98.9 | 99.6 |
| Building (Light) | 98.9 | 99.2 |
| Building (Med) | 94.8 | 99.8 |
| River Network | 99.9 | 99.9 |
| Road Network | 97.0 | 99.1 |
| Coastline/Shoreline | 93.0 | 97.1 |
| Built-up Area | 98.5 | 99.2 |
| Airport | 99.6 | 99.6 |
| Port Facility | 87.8 | 97.7 |
Table 3. Percent accuracy after multiple passes on the panchromatic imagery.
| Feature | Pass 1 Accuracy | Pass 4 Accuracy |
| Building (Dark) | 91.7 | 95.3 |
| Building (Light) | 89.9 | 97.1 |
| Building (Med) | 75.8 | 92.1 |
| River Network | 99.0 | 99.6 |
| Ditch | 98.9 | 99.7 |
| Built-up Area | 88.5 | 89.4 |
Table 4 shows that the Feature Analyst can reduce the labor rate over hand digitization by an approximate factor of over 150 times. Labor hour estimates for each feature on a 1-degree cell. M signifies multispectral while P indicates panchromatic.
| Feature | Head-up Digitizing | FA Labor Estimate | Labor Saving Ratio |
| Airport M | 2 | 0.1 | 20.0 |
| Buildings (all) M | 650 | 2 | 325.0 |
| Building (all) P | 775 | 4 | 193.8 |
| Built-up Area M | 10 | 1 | 10.0 |
| Built-up Area P | 5 | 1 | 5.0 |
| Coastline/Shoreline M | 250 | 1 | 250.0 |
| Ditch P | 250 | 2 | 125.0 |
| Port Facility M | 50 | 1 | 50.0 |
| River Netowrk M | 75 | 0.5 | 150.0 |
| River Network P | 225 | 1 | 225.0 |
| Road Network M | 450 | 3 | 150.0 |
| Total | 2,742.0 | 16.6 | 165.18 |
In this article, we investigated the utility of a new paradigm in feature extraction software, the Feature Analyst. Traditional techniques for feature extraction are slow, laborious, expensive, and inadequate. The Feature Analyst uses machine learning to assist and automate the process of feature extraction. The software interface is simple enough to be fully leveraged by a GIS technician, yet results from this article clearly show that the Feature Analyst (a) reduces the labor time of heads-up digitizing by several orders of magnitude, and (b) mitigates almost all clutter with its hierarchical learning technique.
This work was supported by NIMA contract NMA201-01-C-0016 and National Science Foundation grant IRI-9734419.
Mitchell, T., Machine Learning, Boston: MIT Press, 1997.
Opitz, D. 2002. "The Use of Spatial Context in Image Understanding." Ninth Biennial Remote Sensing Applications Conference.
Oswald et al. 1998. "Cost-Benefit Analysis for Geographic Information System Implementation." New York State GIS Coordinating Body report.
Palhang, M., A. Sowmya, (1997). Two Stage Learning, Two Stage Recognition, Proceedings of Australian Joint Conference on Artificial Intelligence, Perth Australia, 2-4 Dec. 1997.
Quinlan, J.R., Induction of Decision Trees, Machine Learning, (1):, 1986, 81-106.