New opportunities to leverage image data and point clouds for multiple applications are increasingly feasible as high-resolution remotely sensed imagery and LiDAR become accessible to a wider group of users. Many different types of data are available from diverse sources, while hardware and software capable of processing very large datasets have improved the accuracy and speed of analysis. By utilizing artificial intelligence and machine learning to facilitate feature extraction, more useful information can
be generated to add value to a GIS.
Feature extraction is very useful in many areas of GIS analysis, including agriculture, environment, forestry, disaster response and state and local government mapping. Results quickly identify areas of stressed vegetation, possibly from drought or disease, or calculate the quantity of a specific type of tree. Emergency responders may use feature extraction to locate all the buildings in an area that has been damaged by an earthquake, hurricane or tornado, to more quickly get help to victims, and governments find land-cover maps ideal for analyzing stormwater runoff from impermeable surfaces, unpermitted buildings, open space and parks, transportation networks and more.
Automated feature extraction is the holy grail in the mapping community. Increasing productivity by reducing time-consuming manual extraction processes is key. Pixel-based image classification was a step in the right direction. This approach is based exclusively on the digital number of each individual pixel, so the resulting classification considers only spectral information. As each pixel is analyzed individually, identifying discrete objects is difficult. The increased use of very high-resolution data which clearly shows distinct objects makes pixel analysis less attractive, as it loses valuable intelligence throughout the process.
More recently object-oriented classification is seen as an alternative technique that takes advantage of the wealth of information that can be discovered in high-resolution data. By using a set of hierarchical rules, this type of analysis incorporates a broad range of object features, such as spectral values, shape and texture, and combines these with machine learning techniques, independent of the data source (images or point clouds). By creating objects based on spectral and spatial homogeneity definitions, followed by guided object classification according to pre-loaded or user-defined parameters, the result is more accurate and achieved with less human intervention.
The object-oriented image classification workflow is quite straightforward. The user starts by loading data into the analysis software, creating objects (segments) of homogenous areas using automated / guided workflows, then classifying within the segments using knowledge based or machine learning classification techniques. Adjustments are made for factors that may affect the analysis, such as resolution and accuracy of the image; then the analysis results are exported into a GIS.
When the goal of image analysis is to identify and measure homogenous image features, Object-Based Image Analysis (OBIA) is superior to pixel-based methods. Automated functions and batch processing expedite the process, and even more importantly, machine learning continues to improve as more intelligence is gathered and the knowledge base grows.
eCognition is an out-of-the box object-oriented image analysis software package ideal for forestry, agriculture, environment, disaster response and land-cover mapping. For more information about eCognition’s feature extraction and image analysis capabilities, go to https://geospatial.trimble.com/what-is-ecognition