Title

Vicmap Vegetation - Tree Density Polygon 2001 

Alternate title

TREE_DENSITY_2001 

Purpose

TREE_DENSITY is intended to show both the presence and density of tree cover at 1:25,000 scale. The potential uses are many, however, the initial purpose is to provide a detailed, comprehensive and consistent statewide dataset from which tree cover monitoring can be based. TREE_DENSITY provides the most detailed statewide picture of tree cover available to date and is an excellent source of data for any applications requiring the identification of small patches of remnant tree cover such as connectivity analyses and habitat modelling. The scattered density category provides a first in the representation of scattered woodlands and sparse tree cover. The dataset also lends itself well to map presentation and makes an excellent backdrop for other thematic information shown on maps. 

Credit

Cat Gilbert,John White. 

Supplemental Information

History: CNN machine learning was used to assign each pixel in aerial photography as tree or not tree. Relationship to other Datasets: Although independent from other CGDL tree cover datasets and providing greater detail, TREE_DENSITY has some common themes with the following datasets: TCH100_9095, TCH100_9093 and TREE100. Each, in some way represents tree cover and each was derived from satellite imagery. Each also uses the same definition for tree cover. Current Design Issues: TREE_DENSITY is derived from SPOT Panchromatic imagery. Each image covers an area of 60x60km and approximately 90 images are required to cover the State of Victoria. As a result, TREE_DENSITY has been derived from many images of different dates. This not only presents an issue for its primary purpose as a monitoring base, but also for layer design. It is theoretically possible for polygons within any map section to be derived from images dated as far apart as 10 years. It is even common for individual polygons to have been derived from images of at least two different dates. It is therefor imperative to carry date information as a polygon attribute. Several options are available to deal with date information. Firstly, the polygon boundaries of the images used to derive the layer can be brought into the dataset. This would divide some tree boundaries by an artificial break and adversely affect any analysis of the data by polygon area and density class. To avoid splitting tree density polygons by date, it is neater to carry multiple date attributes that can be used to define the upper and lower dates of the images used to derive each polygon. Where polygons were derived from one image, both upper and lower dates will be the same. Where more than two images were used to derive a polygon, the earliest date and the latest date are used. The date attributes are also filled in for non tree polygons. This provides the additional information that there were no trees at the given date(s). Future Design Issues: A numeric code for density class may be implemented as the current character code may prove to be cumbersome. A new raster TREE25 dataset may also be implemented representing tree cover before it is grouped into density classes. Related Documents: None An Arc/Info coverage for each AMG zone contains details of all SPOT Panchromatic images used, including image date, image name and extent of image used. These coverages are maintained in project: p_t25/covers and are named meta-index54 and meta-index55. Further image details can be obtained from the Image Archive Access Database maintained by LIG. 

Status

Completed  

Horizontal Accuracy

25m  

Code

4283 

Maintenance and update frequency

Irregular  

Topic category

• Environment
• Biota
• Imagery base maps earth cover
• Farming

Use limitation

Data is available under a Creative Commons licence. 

Classification

Unclassified  

Statement

Dataset Source: 20cm high resolution aerial photography. Dataset Originality: Derived 

Description

Collection Method: Digital classification 

Description

This data was is derived machine learning output. The tree extent dataset was created by pixel-by-pixel classification of DELWP¿s aerial photography into two classes: tree or not tree. The tree extent data was created using a machine learning method called semantic segmentation. In semantic segmentation, the machine learning model is trained on aerial images, learning from a corresponding raster mask that indicates which pixels represent trees. From the training examples, the machine learning model learns to distinguish pixels that belong to woody vegetation from those that belong to all other features, such as ground cover, roads, and human-built structures. The model was trained on 20cm aerial photography. After processing the state-wide tree extent a waterbody and crop mask were applied to remove any incorrect areas classified as tree cover found over those areas. No additional human intervention post processing was performed on the data. 

Classification

Unclassified  

Attribute Quality

Comments

The three density classes are derived by automated grid cell processing, thereby ensuring consistency. The attribute accuracy of the density classes is solely dependent on the accuracy of the SPOT Panchromatic image classifications. The "Level" polygon attribute has been added as an indicator of the classification accuracy. 

Positional Accuracy

Comments

The positional accuracy, determined by the geometric rectification of the source SPOT Panchromatic images, and reported as Root Mean Square Error, is up to 1.5 pixels. As the images have 10 meter pixel resolution, this translates to 15 meters. 

Conceptual Consistency

Comments

All polygons are automatically generated in a raster to vector conversion. All polygons are closed and labelled consistently. All relationships between attributes are logical. 

Missing Data

Comments

A total of 592 map sections for zone 54 have been completed to date. A total of 203 map sections for zone 55 have been completed to date. Completeness Verification: Three Levels of verification have been described for TREE_DENSITY and the Level is carried as a polygon attribute. The tree cover classifications of the SPOT imagery have, in general, overestimated the occurrence of trees. Although spatially very accurate, the classifications need to be "cleaned" by a process of visual interpretation and manual editing. The notion of Levels of data has been introduced as an indicator to the data user of how much manual editing and/or field verification has been undertaken on the original tree cover classification. Its purpose is to serve as a qualification of the data. The Levels describe an evolutionary "cleaning" path, moving from an excellent starting picture (Level 1) through to the best picture achievable from the source imagery (Level 3). Level 1 data : Tree cover data has been partially (on screen) edited for obvious classification errors. Further editing can be undertaken to remove misclassifications still present in the data. The majority of the remaining errors are water bodies, wet areas and townships misclassified as tree cover. No field checking has been undertaken. Level 2 data : Tree cover has been (on screen) edited as far as practically possible to remove misclassifications and known errors. Further visual interpretation and on screen editing would not significantly impove the data. No field checking has been undertaken. Level 3 data : Tree cover has been (on screen) edited as far as practically possible to remove known errors and misclassifications, and has been checked in the field with additional editing undertaken based on field check results. 

Excess Data

Comments

TREE_DENSITY is a second generation derived dataset. Three density classes have been derived from a classification of SPOT Panchromatic imagery, which itself is a derived dataset. Moving from Dense through to Scattered, the TREE_DENSITY boundaries are, in effect, more generalised or stylised. Dense tree cover boundaries will be tangible, physical edges of patches of dense trees and will be observable on ground. While Scattered tree cover boundaries will not necessarily be physically obvious at ground level. The Dense class : represents tree cover of approximately 80+% density. It has a minimum patch size of 5 hectares (smaller patches will be medium class). And it allows for minimum gaps in tree cover of 0.1 hectares (smaller gaps are closed over). The Medium class : represents tree cover of approximately 50-80% density. It has a minimum patch size of 1 hectare (smaller patches may fall into scattered class). And it allows for minimum gaps in tree cover of 0.25 hectares (smaller gaps are closed over). The Scattered Class : represents tree cover of approximately 10-50% density. It has a minimum patch size of 1 hectare (smaller patches are left out). And it allows for a minimum gap of 0.1 hectares