Using High Probability Image Segmentation to Create Tree Canopy Regions for Zonal Change Detection

For populated areas subject to tree ordinances, the urban forester needs to know how much tree canopy is lost when new buildings and residential areas are built, and then decide how much and where new tree canopy should to be planted elsewhere in their jurisdiction. This is important to help meet the requirements of the local tree ordinance and maintain the balance of man-made structures with the natural environment. Zonal Change Detection is a process that urban foresters can use to gain the information necessary to make these decisions (Figure 1).


Figure 1: Obvious areas of recent tree canopy removal as indicated by 4-band NAIP imagery.

Created initially with the property appraiser in mind, the Hexagon Geospatial Zonal Change Detection workflow in ERDAS IMAGINE uses two dates of imagery combined with vector polygon data, also referred to as zones, to detect change at the parcel level (Figure 2).


Figure 2: ERDAS IMAGINE Zonal Change Detection for Property Appraisal

The workflow utilizes the robust processing power of the ERDAS IMAGINE Spatial Modeler behind a simplified user interface (UI) designed for non-expert users of remotely sensed data. The simplified UI uses the concept of “My ERDAS IMAGINE” which allows any user to customize the ERDAS IMAGINE Layout and Ribbon interface to meet their workflow needs. The “My ERDAS IMAGINE” customization can be applied to any vertical markets such as forestry, agriculture and urban planning.

In the case of the urban forester we have 4-band 2009 NAIP scene and 2010 NAIP scene for an urban environment and are interested in tree canopy loss. However, we do not have appropriate polygon data that can be used as zones, so we have to create our own. In addition, we want the zones to delineate the natural boundary of the tree canopy structure and not man-made political boundaries that are often the case with as-built structures. A sub-workflow inside of IMAGINE Objective is used to accomplish this task.

IMAGINE Objective is a powerhouse collection of classifiers and operators designed specifically to extract features such as roof tops, impervious surfaces, road centerlines, tree canopy and other challenging features of interest from high-resolution geospatial data. In this case we are only using a sub-workflow that only uses Single Feature Probability, FLS Image Segmentation and Raster to Vector conversion operators from IMAGINE Objective. We do not need to execute the Object-based classification and Vector Clean-up operators.

The process starts off with the collection of a few training samples from the 2009 NAIP image that represent tree canopy. We apply the training and the Single Feature Probability pixel cue to identify high probability tree canopy pixels. Then we apply the FLS Image Segmentation to delineate natural boundaries of the tree canopy based on spectral, texture, shape and size settings and a target size image segment. During the image segmentation process, the high probability values from the previous step are written into the attribute table for each segment as a probability value ranging from 0 to 1 (Figure 3).


Figure 3: IMAGINE Objective Single Feature Probability tree canopy pixels extracted from NAIP imagery.

At this point we could isolate the tree canopy pixels in IMAGINE Objective, however we are going to skip that step to convert the segments to vector polygons and then use the attribute selection process in ERDAS IMAGINE Zonal Change Detection to focus our change detection analysis on only tree canopy image segments using the range of probability values generated by IMAGINE Objective (Figure 4).


Figure 4: Vectorized image segments extracted from NAIP using ERDAS IMAGINE FLS Image Segmentation. Each image segment contains a probability value on whether it is tree canopy or not.

Both the before and after 4-band NAIP image and the image segments zones are input into the ERDAS IMAGINE Zonal Change Detection workflow. We adjust the sensitivity settings based on anticipated pixels value differences of the same objects in the before and after image, the targeted spatial detail of changed features, and the amount of shadow visible in the before and after image. Then we run the change detection process only within image segment zones that have a high probability of being tree canopy. In this case we chose 80%, or 0.8 probability, as the threshold using the attribute inquiry filter to determine the region of analysis. By applying the change detection to only this region of zones, we are reducing the possibility of false positives related to other vegetative features, in addition we are reducing unnecessary processing time for larger datasets.


Figure 5: Tree canopy image segment zones that have changed because of removal of tree canopy.

Reviewing the results from ERDAS IMAGINE Zonal Change Detection clearly identifies the 2009 NAIP image segments (zones) where tree canopy has been removed in the 2010 NAIP image (Figure 5 & 6). Furthermore, there is a sharp cut-off in the sorting of the zones where there is not change. Those unchanged image segment derived zones can be saved for a future project.


Figure 6: Tree canopy image segment zones that have changed because of removal of tree canopy.

Utilizing this approach to create natural zones from image segmentation and limiting the amount of manual review using Zonal Change Detection can save considerable time to determine how much tree canopy is removed from year to year. For populated regions subject to tree ordinances, an urban forester or tree canopy expert can determine how much tree canopy is lost and then decide how much and where tree canopy is required to be planted elsewhere in the jurisdiction to meet the requirements of the local tree ordinance to maintain the balance of man-made structure and tree canopy.