Skip to content Skip to sidebar Skip to footer

Urban tree cover provides not only a valuable cooling effect in the congested concrete jungles that are swallowing up our landscapes and lives as an ever increasing number of people flock to urban areas throughout the world, but they also generate social and economic benefits as they improve the livability of cities.

A number of national and local movements have been launched in the past to revitalize urban trees, green spaces and even establish urban forests. As we commit to these fragile environments it becomes increasingly important to monitor them. Remote Sensing offers an effective tool for mapping and monitoring such spaces over large spatial and temporal distances as field-based surveys are often time consuming and costly.

A recent article by Shirisa Timilsina et al., entitled “Mapping Urban Tree Cover Changes Using Object-Based Convolutional Neural Network (OB-CNN)” published in MDPI’s online journal remote sensing, examines the use of combining two different, yet complimentary, remote sensing methodologies to identify urban tree cover over a 10 year period in a suburban environment in Hobart, Tasmania, Australia.

The combination of Object-based Image Analysis (OBIA) with Deep Learning techniques, such as Convolutional Neural Networks (CNNs), is a unique strength of the Trimble eCognition software; a strength we call method fusion.

For this study, the authors used a combination of different data types to account for the spatial and temporal requirements. 4-band QuickBird satellite imagery (60 cm) was available for the 2005 data set. The 2015/16 data set was made up of a combination of QuickBird and 3-band Google Earth imagery. In addition, LiDAR point clouds were acquired in 2008 and 2011 available and used to create the canopy height models (CHMs). 

After the input data sets were prepared for ingestion into eCognition, the CNN model was developed. First, training samples representing two land cover classes, tree and non-tree (other) were created. When creating sample patches for CNN it is important to include variation as this will prevent the model from learning only limited circumstances in the data –  “The tree class represented urban trees of different species within the sample patches, and the other class represented all other nontree features, including grassland, bare land, buildings, water bodies and roads”.

As anyone who has experimented with Deep Learning knows, one of the most tedious and time consuming tasks is generating a good sample data set. But, eCognition includes wonderful tools that can help automate this process. In this case, the authors chose to use traditional OBIA techniques to generate samples by initially segmenting and classifying image objects based on their two CNN training classes. This eliminated the need for manual labeling. According to the Timilsina et al., it took only “five minutes to generate the sample patches for each class”!

Unfortunately, the computer does not [yet] take over all the decision making when it comes to designing the CNN architecture – some parameters still require research, experimentation and trial & error. In this research “8000 sample patches were generated for the tree and other classes, separately” and a sample patch size of 22 x 22 pixels was used. “The selection of sample patch size was done by trial-and-error approaches. Values smaller than 22×22 increased tree canopy detection error, whereas values larger than 22×22 missed some of the small trees. Most of small trees in the study area were found to be within 22×22 pixels”. All available spectral bands were used when creating the samples: RGB+NIR for the 2005 data set and RGB for the 2015/16 data set.

Example of 22 x 22 pixel sample patches generated from the CNN – the top row shows samples for the class tree and the bottom row shows samples for the class other (nontree).

The authors chose to use a simple model architecture and used one hidden layer and determined that a learning rate  of 0.0015 was best based on testing various scenarios. The training phase took about 30 minutes to complete – fast processing can be achieved if samples are stored in the default raw format.

Subsequently the model was applied and heatmaps were generated for the different classes and time periods. The resulting heatmaps were there smoothed using a gaussian filter and the heatmap for the class trees was then put through a morphology filter.

Results of segmentation based on original input image compared to those achieved with inclusion of heatmap layer.

The heatmaps were then used as input for the multiresolution segmentation and objects with a CNN probability >50% were classified as trees. To reduce noise in the classification results, the authors decided to use OBIA-logic to remove objects with similar spectral characteristics to trees by applying CHM and NDVI thresholds. In addition, image object refinements were run using pixel-based object resizing, remove objects (to establish a minimum mapping unit) and merge region.

The results were then exported as a shapefile and an accuracy assessment was performed to calculate Precision, Recall, F1 measure and Intersection Over Union, whereby:

  • Precision is used to answer the question “How many of the classified pixels are trees?”
  • Recall is used to determine the proportion of actual tree pixels that were classified as trees in the image.
  • F1 measurement is then the balance between Precision and Recall.
  • Intersection Over Union (IOU) measures the accuracy of the classification results based on ground truth (an IOU of 100% means exact overlap with ground truth pixels, 0% means no overlap).

Within this study, the mean IOU was %70 and the mean Precision and Recall values were 87% and 85% respectively. The F1 measure ranged from 77-94%. Finally, an overall accuracy of 96% was calculated for the 2005 data set and 98% for the 2015/16 data set.

When the change detection was applied to the different time periods, the authors determined that there was a net tree cover loss in all areas of the study region – “The highest losses of tree cover between 2005 and 2015 were in those areas where new developments of houses occurred amongst indigenous trees. The removal of older local indigenous trees tends to occur gradually, as they drop limbs. The older suburbs and those developed on farmlands did not exhibit high levels of net tree losses”. 

I hope that these types of study will motivate local governments to adjust policies, especially in areas slated for urban expansion. Although it is not possible to prevent all tree loss in such areas, it would be advantageous to see plans include the value of urban trees and forest as they improve the quality of life in such areas.

Follow Us on Social Media

For more news on eCognition