A Fast, Multi-class Algorithm for Classifying Aerial Images.

J. W. Gates, Ren Togo, Hideo KitajimaThe Journal of The Institute of Image Information and Television Engineers 2002 · DOI

Abstract

An approach is described for classifying images that does not require first dividing an image into small blocks and then classifying it based on the features of the individual blocks, as do traditional approaches. Instead the line features of the entire image are extracted and used, along with the pixel intensity, to classify each pixel in the image. To increase the classification speed, a fast line-extraction algorithm has been developed that extracts the line features directly from the original image without pre-processing. A classification tree with single variable splits is used to classify the image. Testing of a five-class aerial-image classification algorithm showed that it had an average error rate of 17.6%. Running on a 600-MHz Pentium III processor, it had an average classification time of 2.18 seconds for 512×512 grayscale images. This approach can be used for many different applications by training the classification tree with the desired classes.

Introduction

In the early 2000s, the demand for large-scale environmental monitoring—spurred by international initiatives like the Kyoto Accord—required the rapid processing of millions of satellite and aerial images. Traditional classification methods typically relied on "block-based" approaches, where an image is partitioned into a grid and each block is classified independently. This methodology suffers from two primary drawbacks: feature fragmentation, where grid lines bisect critical objects, and high computational overhead due to intensive pre-processing steps like Sobel filtering and erosion.

This paper proposes a "global" classification method that extracts line features from the entire image context, avoiding grid-induced errors and significantly reducing processing time by eliminating traditional edge-detection pipelines.

Methodology

Fast Line Extraction

The core innovation is a line-extraction algorithm that operates directly on raw grayscale data. By scanning for intensity contrasts exceeding 30 units and tracing the perimeters of uniform regions, the system uses simple linear regression to fit line segments. This bypasses the computationally expensive pre-processing steps typically required for edge detection.

Feature Distribution

Following extraction, line properties such as length, slope, and distance are "distributed" to neighboring pixels. For each pixel, the algorithm identifies the five nearest lines. This "push" method is significantly more efficient than a sliding-window "pull" approach.

Figure 1: Distribution of distance features for Forest and Water classes, showing clear separation based on pixel distance to lines.

Classification Tree

A decision tree with single-variable splits is employed for final classification. This choice ensures high execution speed and provides robustness against noise in manually labeled training data. The algorithm uses 17 additional arrays to store feature data for every pixel, allowing for near-instantaneous processing at the cost of higher memory usage.

Experimental Setup and Datasets

The algorithm was trained and tested using a variety of 8-bit grayscale aerial images. The training set included four images representing diverse landscapes, while the test set comprised three 512x512 images containing five target classes: fields/beaches, buildings/roads, water, forests, and shadows.

Figure 3: Examples of aerial images used in the study, including urban, water, suburb, and forest environments.

Results and Discussion

The system demonstrated remarkable efficiency, classifying a 512x512 image in an average of 2.18 seconds on a 600-MHz Pentium III processor. This represents a nearly tenfold speed improvement over contemporary neural network approaches of the time. The average error rate across the five classes was 17.6%.

Analysis of the processing time revealed that the line extraction itself is extremely fast (0.09s), shifting the primary bottleneck to the distribution of features to neighboring pixels, which accounted for approximately 80% of the total runtime.

Conclusion

The proposed global feature extraction and fast decision tree classifier provide a high-speed alternative to traditional block-based aerial image analysis. While some errors occur at class transitions (e.g., beach-water boundaries), the algorithm's adaptability and speed make it suitable for large-scale environmental monitoring. Future work will focus on refining the classification tree to mitigate boundary errors and optimizing memory usage for more constrained hardware environments.

Summary updated

Abstract

An approach is described for classifying images that does not require first dividing an image into small blocks and then classifying it based on the features of the individual blocks, as do traditional approaches. Instead the line features of the entire image are extracted and used, along with the pixel intensity, to classify each pixel in the image. To increase the classification speed, a fast line-extraction algorithm has been developed that extracts the line features directly from the original image without pre-processing. A classification tree with single variable splits is used to classify the image. Testing of a five-class aerial-image classification algorithm showed that it had an average error rate of 17.6%. Running on a 600-MHz Pentium III processor, it had an average classification time of 2.18 seconds for 512×512 grayscale images. This approach can be used for many different applications by training the classification tree with the desired classes.
BibTeX
@article{gates2002fast,
  author = {J. W. Gates and Ren Togo and Hideo Kitajima},
  title = {{A Fast, Multi-class Algorithm for Classifying Aerial Images.}},
  journal = {The Journal of The Institute of Image Information and Television Engineers},
  year = {2002},
  doi = {10.3169/itej.56.826},
}
Figure