Embedded Files
Determining Land Coverage Using Aerial Imagery and Image Classification
Problem:
Black Water National Wildlife Refuge Managers are developing a management plan for the refuge in Maryland. To do this they need to develop a methodology for classifying land coverage data into six key land classes; Water, Wetlands, Forest, Cultivated Field, Barren and Developed/Impervious Surfaces. Aerial imagery for a portion of the refuge has been provided for development of the methodology.
Analysis Procedures
ArcGIS Pro was used to analyze the aerial imagery provided by refuge managers in both true color and infrared. I created six classifications in the Training Samples Manager and then created sampling polygons for each of the classifications. The Classify tool then used these polygons to identify other spectrally similar areas. I examined the resulting raster image to identify errors in classification. Where errors occurred, I created additional sampling polygons to better “teach” the software how to classify these areas and reran the Classify tool. After each time I ran the Classify tool, I exported the raster and used the Build Raster Attribute Table tool to gain a count of pixels in each category. With this information I was able to add fields to the attribute table and calculate the square footage and acreage of each classification.
The four-band imagery (R,G,B,IR at 0.3-meter resolution) provided by the Refuge Managers was added to the analysis twice. One layer was configured to a true color image and the other to a false color infrared image to make for easy toggling between images for identifying classes. I created a new schema in the Training Samples manager consisting of the six categories identified by the Refuge Managers and created four training samples for each of the six categories. I then used these samples to run the Classify tool which returned a raster image where each pixel was identified as one of the six categories. I exported this image as a .tif type image and ran the Build Attribute Table providing an added field of “Count” containing the number of pixels in each category. In the attribute table I added two fields and configured them in the Calculate Field tool to determine square footage and acreage of each category. Examining the image, I found areas misclassified in four categories such as water classified as impermeable surface. These misclassified areas indicated opportunities for refining the classification.
To further clarify the image, I went back to the Training Samples manager for a second round of analysis. I added training sample polygons to all classifications with the exception of Wetlands and Developed/Impermeable Surfaces. Several of these added polygons were very small to help the Classify tool differentiate between coverage in areas where trees shaded terrain. Once again, I ran the Classify tool and processed the image again to add an attribute table that calculated square footage and acreage for each category.
Results
Click images for a larger version
Figure 1. First round of training samples created for classifying land coverage types over the 484 acre section of the refuge.
Figure 2. Results from first round of classification. Notice areas of the river misclassified as "Developed". Also areas where trees are casting shadows tend to be misclassified as well.
Figure 3. More training samples were added to further refine the spectral signatures of misclassified areas and then the analysis was run a second time.
Figure 4. Misclassified areas were fixed in the second round resulting in a 22.6% increase in water surface while developed surface decreased by more than 85%. Misclassified crop areas took forest acreage from 224 acres in the first image to just over 164 acres.
Application and Reflection
Problem description:
Supervised image classification is one way to monitor the clearing and regrowth of forests. Illegal clearing of rainforests is a continued problem in many parts of the world. This would have applications for monitoring rainforest loss/reclamation or managing stands of trees and predicting future timber harvests based on the rate of change in subsequent images.
Data needed:
By using high resolution satellite imagery such as Sentinel 2 imagery or preferably aerial imagery (R,G,B,IR) it would be possible to monitor the regrowth of forests over a much larger area.
Analysis procedures:
The spectral signature of bare ground is obviously drastically different from fully forested areas, but I realized in doing this activity that the spectral signatures vary in terms of the density and indirectly the age of the vegetation. I experimented with this using the image provided in this activity where I noticed several fields in the southern part of the image behind the house that appear to be former farm fields with some trees growing. I confirmed this by looking at more recent satellite imagery which showed continued growth in these areas undergoing succession. I built a new schema for the image and classified the area into “No Growth/Bare Soil”, “Crops/New Growth”, “Medium Growth”, “Full Forest” and “Water”. The results showed that the area where the former farm field was undergoing succession had an even mix of categories ranging from bare soil to medium growth whereas the forested areas had scattered spots of medium growth.
Google Sites
Report abuse