Logical classifier for Fung Shui (native) Woodland in Hong Kong
Remote sensing and Forestry tracing back to 1924
Fung Shui Wood was the native forest (woodlands) in Hong Kong, they were well preserved due to the remote locations, rugged terrain and local beliefs. In Hong Kong, we can trace our aerial image records dating back to 1924 and those black and white aerial images already help us for the demarcation and differentiate the woodland, grassland and shrubs. Then, with the pre-WWII and post-war forest change detection, we can determine the native woodlands as preserved by the surrounding villagers. For example, Lamma island and She Shan and Tai Om are the typical remaining native (fung shui) woodlands in Hong Kong.
Simple logical classifier for the native forest detection
For the classification of the native forest, the aerial images and land lots (derived from the lots index plan) are the major input in this logical classifier. Aerial images serve as the major way to determine the vegetation cover and land lots to correlate with the associations between the old villages to the vegetation covers. We will classify the vegetation cover types first, then apply the buffering with the land lots to determine the woodlands near the villages.
After that, by comparing the 1924 and 1963 aerial images differences and similarities. The native fung shui forest and secondary forest can be classified. The loop will be if both 1924 and 1963 vegetation classification are woodlands, it will be the Fung Shui Wood. Otherwise, there are difference between 1924 and 1963 images, then we need to further identify the differences into two conditions. First, it was non-vegetation in 1924, but it was woodland in 1963, so it would be considered an “artificially” generated secondary forest. Otherwise, if it was vegetation in 1924 and woodland in 1963, it would be considered “regenerated” forest.
Remote sensing with artificial intelligence
In the past, processing old black-and-white images was a challenging task, with human interpretation and evaluation taking a lot of time and effort during the vegetation classification stage. Now, many pre-classified artificial intelligence models such as Meta Segment-Anything, DeepAI and Stability.ai can help map vegetation cover in a faster and more accurate way.
Validations and evaulations
Since there is no official public data for Fung Shui Forest (native forest land) verification, we can only refer to PlansD SSSI boundaries. The results cross-referenced with SSSI boundaries are approximately 55.9% and 66.2%.
Judging from the classification of the logical classifier, the accuracy is not high. However, classified native forest areas are still valuable for forestry purposes. Since the setting of SSSI boundaries is not the “real” change detection based on historical images, in my opinion, the accuracy is acceptable.
Further development
From outside to inside, normally optical remote sensing can only derive the chemical and general composition from the imaging. Therefore, it is time to discover our forest with various methods, such as active and close-ranging remote sensing.
Active remote sensing different from optical remote sensing in that can work without illumination (light source) and capture not only the surface but also the interior of the ground. LiDAR (Light Detection and Ranging) and SAR (Synthetic Aperture Radar) are some examples of active remote sensing.
In LiDAR industry, they have different capture mediums, including but not limited to spaceborne, airborne, close-ranging and terrestrial. Regarding the concept of LiDAR, the airborne and spaceborne sensor is better to know the canopy structure and density that derive the information from the light beam interacting with woodlands. Besides, handheld, close-ranging and terrestrial LiDAR are for the tree structure analysis. Leaf-On and Leaf-Off data can be derive from those point clouds.
For close-range remote sensing, panoramic imagery and hemispheric photos, help us define and derive ground truth from our field work. Using reprojection techniques and computer vision, we can classify canopy cover, tree numbers, and tree species from field observations.
Disclaimer and Acknowledgement
This is a volunteering pilot study conducted by the Forestree team, specializing in Remote Sensing and Forestry. The study focuses on close-ranging photogrammetry, image processing, and computer vision. The figures in this study were generated by Yu Kai Him Otto, using data obtained from CEDD Airborne LiDAR and LandsD aerial images (HKMS2.0). All data analysis and processing were performed using GIS and Python scripts developed by Otto. All rights reserved.
