This first entry deals with several concepts I have been toying with over the past week, namely the role and origins of “radical thought” in communities, particularly under the lens of governmentality. This discussion is in response to one very central concern I have with my dissertation — what are all the different reasons intentional communities provide an interesting and important framework for analyzing sustainability policy?

Radical Rural Roots

I led a reading group for the NSEP MSc students in weeks one and two on governmentality, technological zones, and global assemblages. I starting thinking about how intentional communities may act as particular zones of qualification that exist at a level between that of the individual and the state. They depend on knowledge produced by or associated with other communities (i.e. appropriate technologies, legal precedence), sometimes preferring it to outside knowledge. Based on my interviews alone, this would appear to be a stronger trend in more radical communities. For example, one member at the Secret Garden Community (SGC) expressed frustration that the community as a whole was wary of experimentation based on inter-community ‘expert knowledge.’ Similarly, when asked how and from what sources knowledge was produced and dispersed, many members’ first responses were the internet or books. At the Hundred Acre Woodland Community (HAWC), however, the experiential knowledge of each community member was equally evaluated and valued at a woodland management meeting I was permitted to attend. That said, at SGC I observed people regularly sharing experiential knowledge informally while working in the garden or milk kitchen.

I would like to further explore the role and origins of ‘radicality’ in communities. I’ve begun reading several articles about the “radical rural” and alienation of intentional communities. I think this ties into some of the points Elizabeth Dunn makes in her chapter on “Standards and Person-Making in East Central Europe.” She describes how standards can marginalize some groups and that these groups find alternative (often illegal) ways around the standard. She posits that this is notably the case in Poland where previously-existing black market infrastructure provides an easy way around EU standards. This connects to intentional communities in considering Shenker’s assertion that societal alienation is a driving force behind the formation of intentional communities (1986).  I hypothesize that increasing alienation and marginalization of people and communities (intentional or otherwise) would correspond both to actions increasingly outside of set societal “standards” and the perception of increasing radicality.  For example, back-to-the-land initiatives like SGC were often in response to a feeling of alienation from food-production systems. There are numerous examples where HAWC has been forced to move outside set standards of zoning, planning, forest management, and diverse definitions of “sustainability” in order for residents to persist in their chosen way of life.

So, some key questions for consideration:

1. Does intra and inter-community knowledge production produce unique zones of qualification?
2. Are these trends stronger in more radical environments?
3. Is there a gap (particularly at SGC) between perceived and actual zones of qualification?
4. Can/do the terms “alienation” and “marginalization” refer to the same situations?
5. What exactly makes my study communities alienated or marginalized? What standards do they conform to or rebel against and why?
   
6. How does the study of (marginalized/alienated intentional communities) apply to the study of sustainable development?

A brief response to this final question. There is ongoing debate as to the utility of sustainable development discourse. On the one hand, it provides a common language of balanced economy, society, and environment that may be used among diverse actors. However, this flexibility also may allow any group to co-opt the term in order to justify its own actions. Thus, many call for a standardization of sustainability. But, as described here, standardization causes marginalization and alienation of those outside the decided standards. I believe intentional communities are unique in that that act as boundary organizations in numerous ways. First, they act at a level intermediate to the individual and the state. Second, they are forced to exist in balance between utopian and standardized ideals (such as sustainability). This (forces? allows?) them to function within unique zones of qualification.

That is just an introduction to some of my thoughts on intentional communities as boundary organizations. More on that next time!

CONTENTS (click each to be directed to that section)

I. Reading Summaries

II. Project Update and Data

III. Land Change Modelling Lab

*Note: you can click on the images to see the larger version in a separate window.

I. Reading Summaries

II. Project Proposals

For my assessed essay project, I think the section on biodiversity analysis may have some useful components. I am currently doing some readings about using remote sensing data to identify species, ecosystem richness, and biodiversity. It may be useful to additionally use some of the analyses provided in this lab.

III. Land Change Modelling Lab

6-1 LCM: Projects and Change Analysis

This section deals with analysing observed change over a period of time once land cover has been classified using supervised or unsupervised methods (see Week 4 lab). I used pre-classified land cover data from central Massachusetts from 1985 and 1999. Using simple visual analysis alone, it is difficult to determine how much land has changed and in what ways (Fig. 1). Using the IDRISI module Land Change Modeller (LCM), I created histograms that demonstrate how much land was converted to various other land forms between the two years. Figure 2 shows overall gains and losses of various land covers. I then broke that down to look at three different categories in particular: Large Residential (Fig. 3), Open Land (Fig. 4), and Cropland (Fig. 5). This change can then be spatially represented using change and transition maps (Fig.6, 7).

Primary Themes:

• Mixed and deciduous forests and croplands have experienced the greatest losses in land cover
• Residences of greater that two acres have experienced the greatest gains in land cover

Secondary Themes:

• (Question 6-1-1) Some area from Large Residential (> 2 acres) is lost to Smaller Residential (< 2 acres and multifamily). Most likely this is due to property subdivision, the transformation of large single-owner estates to multiple small private residences.
• Open land gains are coming primarily from losses in cropland. These losses are focused on the central region of the study site (Fig. 8).
• (Question 6-1-2) The changing land covers most contributing to a decrease in open land are mixed and deciduous forests (Fig. 4). This indicates that land classified as ‘open’ in 1985 in 1999 was classified as mixed or deciduous forest. This conversion could be classified as secondary forest regrowth.

Question 6-1-3 Comparing the map of change to large residential to the trend map for cropland, what can you conclude about the main driving forces of change in this area of Massacusetts?

Answer: From those two trend maps alone, there is little I feel I can conclude about the main driving forces of land cover change because they do not directly correlate to each other – one is all categories to large residential, the other is cropland to open land alone. The maps indicate that where cropland has the most significant changes to open land (central study site) corresponds to the area of least change to large residential. One might therefore be tempted to designate these two categories as the driving forces of land change in the entire area. However, the total hectares of land change in the croplands/open lands category is on the scale of half the size of conversions from forests to residential.

We know from Figure 2 that mixed and deciduous forests and croplands have experienced the greatest losses in land cover and residences of greater that two acres have experienced the greatest gains. Then, we know the primary contributor to deforestation is a change to large residential areas. We also know cropland has largely been converted to open land and much open land has been replanted with mixed and deciduous forest. However, the conversion to forest has not been in areas previously cropland (as far as we know from this data). So I would say the driving forces of land change are primarily urbanization of agricultural and forest lands. However, while there seems to be a trend to reforest certain areas of open land, there is not a trend to relocate or re-establish agriculture in this area.

6-2 LCM: Transition Potential Modelling

The next two sections use data from the Bolivian lowlands, Chiquitania, to create land cover change models with different variables and variable combinations to anticipate land cover in the future based on trends seen between 1986 and 1994 (Figure 9). This first section creates transition potential model maps for four different land cover categories. To do this, I used a multi-layer perceptron (MLP) neural network. Unlike a simple linear regression, a neural network allows me to analyse multiple variables at once, much like a multi-variate regression. As opposed to the previous section, all land change was conversion of various land covers into a single category (anthropogenic disturbance). The primary change categories are mapped in Figure 10.

The first step in creating the model was to determine which variables to use to best explain change between 1986-1994. I looked at seven different  (continuous quantitative) variables: distance from previous disturbance, distance from streams, distance from roads, distance from urban areas, elevation, slope, and likelihood of transition. The latter variable was created from the qualitative layer of land cover by recalculating land cover types as probabilities of transition over the given time period (Figure 11). The resulting transition potential model maps are displayed in Figure 12.

6-3 LCM: Change Prediction

This section uses the transition potential maps from section 6-2 to predict land cover in the year 2000 based on trends established between 1986-1994. Models of three different degrees of complexity were performed:

1. Figure 13: Model was run using all static variables, i.e. assumes variables like road infrastructure and location of development remain as they were in 1994 at the start point of the model.
2. Figure 14: Model run with distance from disturbance set as a dynamic variable. This is done by running the model in iterative stages, recalculating the distance from disturbance variable based on the new (anticipated) scenario. This particular model was run in three stages, i.e. once every other year.
3. Figure 15: Model run with distance from disturbance and roads set as dynamic variables, reserve area constraints in place, and in six stages (i.e. one per year). The reserve area constraints are based on indigenous forest reserves that do not allow for increases in anthropogenic disturbance. Therefore, the model does not include them.

*Note: Each model contains both a hard and soft prediction. Hard predictions show areas as they are anticipated to look in the future. Soft predictions give a buffer range of ‘areas of concern’ that have a potential to be affected.

A final set of models was run for a long prediction of thirty years at 1, 2, 4, 8, 16, and 30 stages. It appears that increasing the number of stages used increases the amount and connectivity of anthropogenic disturbance. Figure 16 shows the outcome with one stage versus 30 stages.

6-4 LCM: Validation

In this section I assess the models created in the previous section for error and compare the value of scenario models (hard predictions) with vulnerability models (soft predictions).

I used a map of actual land cover from 2000 to use as a base line for my model results. When assessed on a full-map level, my models had reasonably low error rates (94% accuracy). However, this is because most of the land in the study area remained the same. This does not mean that increases of anthropogenically-disturbed land were not significant. To better assess only anthropogenic disturbance, I did a cross-classification of predicted change to actual change. This produced a map showing areas of correctly-identified change/non-change and additionally areas of false-positive and false-negative error (Fig. 17).

Much of this error can be attributed to a policy change that took place during the mid-1990s. This caused trends in land change to be different from those seen between 1986 and 1994. However, it is important to note that while these changes registered as error in a scenario model, much of the change still fell within areas of concern within the vulnerability model.

Question 6-4-1: Given that there was a major policy change that had a huge impact on land cover change, what can you conclude about the relative benefits of soft prediction? What are the potential drawbacks?

Answer: Soft prediction can better account for unanticipated events because it shows essentially a buffered area of concern rather than a discrete changed landmass. In GIS speak, this means it is more likely that areas of concern will be correctly identified in prediction models. In policy speak, this is important because it allows for the identification of a wider area of concern with graduated degrees of importance. Ranges are easier to work with than absolutes. However, just as is the case with any flexible boundary/situation, a soft prediction leaves more room for dispute which could be a problem if decisions need to be made immediately.

6-5 LCM: Dynamic Road Development

This section uses a portion of the LCM IDRISI module to model anticipated road development. I began by modelling road development with a given set of parameters concerning road length and spacing. I then tried different spacing and growth length parameters to attempt to identify reasonable parameters and assess sensitivity of the model to initial parameters. Below are my findings and an animation of my preferred parameter settings (Fig. 18).

• Increasing length but not spacing narrowed margin of effect around roads but not drastically. Roads looked more realistically spaced to me.
• Increasing spacing but not length did about the same, but more knobbly. Roads were too long without enough branches in this case.
• Doubling both variables seemed to generally increase the area of disturbance, but it also may have just shifted. The roads themselves look too long without enough branches in this scenario.
• Overall, changing the numbers as I did definitely seemed to have an effect on land cover change, but (at least visually) not a drastic one. I found the roads with an increased spacing but not length in comparison to the given figures looked most realistic, but I did not do extensive testing due to time constraints.

6-6 LCM: Habitat Assessment, Change, and Gap Analysis

This section addresses the implications of change by looking at changes in habitat quality and by using gap analysis. Specifically, I looked at implications of land change on bobcats in central Massachusetts. Data for habitat suitability consists of expert assessment of habitat requirements for the species. Land is then identified as primary habitat, secondary habitat, primary potential corridor, secondary potential corridor, or unsuitable. Change in habitat for the bobcat due to land change between 1985 and 1999 can be seen in Figure 19. While all habitat categories experienced some decreases, only suitable primary habitat has experienced a net decrease in area with 44,124 hectares (over 170 sq miles) lost.

I then performed a gap analysis on the 1999 bobcat habitat to look at correlations between protected areas and habitat type in hopes of identifying potential areas still in need of protection (Fig. 20). Although much of the bobcat’s primary habitat is protected, it looks as though over 50% of it still remains unprotected. Also, although much of the secondary habitat is in the eastern portion of the area and possibly of less use (because it does not link up to any primary habitat) there remain some important areas of secondary habitat in the west that, if protected, could provide important range for bobcats that would otherwise be degraded to corridor or unsuitable status.

Note: Interestingly, many of the protected area borders match up with areas of primary habitat while much of the areas of protected secondary habitat seem to be incidentally protected because their borders do not match up. This might be caused by (at least) two things. First, protected areas were created in areas of primary habitat with the explicit purpose of protecting identified primary habitat. Second, the areas may have been protected previously and as development has decreased areas of suitable primary habitat, primary habitat has been restricted to protected areas. More data would be needed to determine which is the primary driver of this phenomenon.

6-7 LCM: Species Range Polygon Refinement and Habitat Suitability

In this section I used range polygons as defined by experts to identify habitat suitability for the species Vicugna vicugna in South America. The initial polygon indicated a discrete area (i.e. suitable or unsuitable) based on state borders. I used six environmental variables to assess land within that polygon for different habitat requirements for the vicugna. The polygon was then refined to show the degree of habitat suitability within the expert-defined polygon (Fig. 21). Then, using the parameters for degree of habitat suitability, I extrapolated the data over a larger region so that habitat range was no longer constrained by state borders (Fig. 22). Interestingly, although the original polygon would indicate that the entire initially-defined area is suitable vicugna habitat, the area actually classified as entirely suitable (value 1.00) in the entire range of the species is a small fraction of the original area.

6-8 LCM: Biodiversity Analysis

In this section I looked at five different biodiversity indicators, alpha diversity, gamma diversity, beta diversity, Sorensen’s dissimilarity index, and range restriction index, to assess amphibian biodiversity of an area in the northern Andes. I looked at gamma diversity and Sorensen’s index using both focal zones (of 50 km) and pre-defined regions.

Question 6-8-1: Using the results [of the focal zone run], how is the region being protected in terms of local richness, regional richness, richness change, species turnover, and protection of endemics?

Answer: In general, the protected areas are in the regions of concern (i.e. exhibiting high biodiversity); however, many key hotspots are missed (Fig. 23/24). The protected regions look as though their borders were decided based on topographical or political issues rather than biodiversity because many protected areas are only moderately biodiverse (in comparison to neighbouring areas) while many of the areas with highest biodiversity are not actually protected at all. However, there appears to be an inverse relationship between basic biodiversity measures like species richness (alpha and gamma shown in Fig. 23) and the ones that account for endemism and regional effects (shown in Fig. 24). This makes sense in an alpine situation – the more rugged the terrain, the more specialised the amphibian species would be. So although the richness may be lower in some of the alpine areas, the species there are more likely to be endemic and highly dependent on a specific environment (and possibly more dependent on protection?)

Question 6-8-2: Compare the results of beta and gamma [from focal zone and regional] runs. How do they compare? What does gamma tell us about the biodiversity of each eco-region? Which eco-region is more diverse? Which one is least diverse?

Answer: The beta and gamma from the two runs identify similar levels of biodiversity in similar areas, showing that the eco-regions generally do match up with different levels of biodiversity (Fig. 25/26)*. Displaying biodiversity by region is helpful for generalising biodiversity in specific areas, but it less accurately represents areas on a micro level. For example, there is a biodiversity hotspot on the west side of the Gamma Focal50 that is not recognised within an eco-region. The Focal50 run is able to identify it while the run by region is not. Overall, the most biodiverse eco-region for amphibians is the tropical and sub-tropical moist broadleaf forests while the least biodiverse is mangroves.

*Note: I think there may be a problem with the beta diversity eco-region assessment because the little spot in the north-east corner is such high biodiversity in relation to the rest of the area that it has thrown off relative measures. Either this is a completely unrecognised biodiversity hotspot, or (more likely) it is a classification error of some kind.

CONTENTS (click each to be directed to that section)

I. Reading Summaries

II. Project Proposals

IIa. Case Study: Forest Governance and Land Cover Change in Intentional Communities

IIb. Actual Geographies of the Radical Rural: The Spatial Story of England’s Intentional Communities

III. Supervised Classification Lab

IV. Unsupervised Classification Lab

*Note: you can click on the images to see the larger version in a separate window.

*For a document version of the lab report, click here.

I. Reading Summaries

Rogan et al (2002) A comparison of methods for monitoring multitemporal vegetation change using Thematic Mapper imagery

Rogan et al compared several techniques to identify land cover changes on a site in southern California. They were particularly interested in finding techniques suitable for identifying land cover change due to urbanization and fires. They first chose five change categories: 1) no change, 2) vegetation increase, 3) vegetation decrease, 4) change in nonvegitated regions, and 5) recharge or flooding of lakes and reservoirs. They then used two techniques to enhance the images for classification and two further techniques to actually classify the land cover. Results were compared to field observations to determine which methods of image enhancement and classification were most successful. Finally, they discussed possible reasons for and implications of the errors of each technique.

I found the examination of error the most interesting part of this paper. I had never thought about using the spatiality of error to help determine its causes. In this case, Rogan et al examined the clustering of errors of commission and omission. Because errors were clustered, not random, Rogan et al were able to identify primary causes of error and suggest ways to decrease that error in the future. I am not sure I understand the difference between image enhancement and classification. I think that during image enhancement the actual assessments of change are undertaken. Then images are actually classified by land cover type in the classification process. However, that may be an oversimplification.

Zhan et al (2002) Detection of land cover changes using MODIS 250 m data

This paper outlines a study in which land cover change identification was attempted using MODIS images with a 250 m resolution and the Vegetative Cover Conversion (VCC) algorithm. Zhan et al report a high accuracy in detecting change, but a low accuracy in attributing any qualitative information (i.e. identification of pixels representing burned ground.) They wanted to identify areas of burn, flood, and deforestation in order to update look-up tables used in the VCC production code. They found VCC detection methods worked better for large-scale detection and that integrating multiple methods of VCC proved the most effective.

I must admit, I felt confused about much of this paper. Firstly, I do not understand the classification of error. They say that only 32% of actual burned pixel areas were detected but that there was 99% accuracy in detecting change. I am wondering if that means it just classified the erroneous 67% as incorrect land cover, but still changing land cover? Also, concerning the flood detection, they reported 90% detection accuracy because they had identified 1966 sq km as having flooded when an actual 1913 sq km were reported to have flooded. This means that they were within 10% of finding the total area, but it does not actually indicate that they identified 90% of the pixels that actually flooded. Using this logic, they could have predicted 1966 sq km of Cambodian flooding in Wyoming and still claimed 90% accuracy! Those were the specific things I saw concerning error classification, but I am still unclear as to how to measure and analyse error. I found this paper very hard to read, so really all I was interested in was the fact that it talked about the fire in Los Alamos that forced my grandparents to evacuate their home!

Combined Essay question

Question: Why is it important to identify, classify, and quantify error in GIS models?

Answer: Any study, particularly a model-based study, is made stronger by identifying sources and types of error. Discussion of error makes the project more replicable and identifies assumptions that must be made in using the study/model in future research and policy. However, identifying, classifying, and quantifying error in GIS models is additionally important for several reasons. From a technical standpoint, assessing for error will assist with future or secondary analyses. For example, a model can be redefined and tweaked to find more accurate parameters. Or, as is the case with Zhan et al (2002), describing error of multiple models may help to identify an integrated system that is “a more robust detection scheme than [could] be provided by a single algorithm” (pg. 349). Also, error investigations themselves can undergo further analysis that can provide valuable data on the nature of the study site and sources of error. Rogan et al (2002) demonstrate an example of this in their spatial autocorrelation analysis of erroneous pixels. This analysis helped identify source of error as shade and terrain-based. The research team can use this information to improve on future models.

From a policy standpoint, identifying error is extremely important. As science-based decision making becomes more expected and integrated into policy, it is imperative that studies be as accurate and specific about (sources of) error as possible. Policy-makers who are less familiar with technical details will tend to take results at face value, particularly if presented in a simple, attractive form (like a map). Although studies with 32% accuracy at detecting burned areas may be able to provide valuable information to the expert world of GIS, it may not be appropriate to use it to replace the Forest Service’s existing detection methods.

II. Project Proposals

I am required by my MPhil to complete one more assessed essay this year based on my elective module (i.e. GIS/Remote Sensing). It is extremely important that I do will on this essay because my two essays last year were somewhat lacking and are bringing down my overall grade for the course. If I hope to get a distinction in my course, I need to get a solid distinction on this paper! Although I anticipate modifying my project over the coming weeks, this is an initial attempt to outline several project ideas. We were advised to come up with three project plans:

A) The sexy, interesting, innovative plan that would be fun and amazing but also risky as far as data collection and interpretation are concerned.

B) A project using the same techniques as Plan A but in a location where data is more accessible (i.e. in Texas instead of Timbuktu).

C) A project based on data rather than a distinct question. In this way, you know the data and the analysis are there, all you have to do is find the right way to frame the question

Being the rebel that I am, I have two project ideas, both of potentially moderate risk, both with a number of pros, cons, and applications to my dissertation research.

IIa. Case Study: Forest Governance and Land Cover Change in Intentional Communities

Background

The Hundred Acre Woodland Community (HAWC)* was established in 2000 by a group of people wanting to live lightly in a community setting. Their missions included veganism, permaculture, minimising or eliminating fossil fuel use, and sustainable forest management. Initial residents purchased a plot of woodland (approximately 32 acres or 129,500 square meters) located in a national park in the south of England. According to the HAWC website, in the 1920s the area was set up for commercial harvest of primarily fast-growing exotic conifers. The area is now divided into three general management areas:

1. Northwest portion: Primarily mature Scot’s Pine
2. Central: Small Ash plantation with an invasion of Sycamore
3. Remaining (largest): Japanese Larch with emergent understory of Hazel, Ash, Sycamore, Rowan, and Oak

One of the goals of HAWC is to manage the woodland to promote the biodiverse broadleaf understory and slowly phase out the monoculture of plantation conifers. They are doing this via the forestry method of continuous cover management. This method focuses on felling select trees or small clearings rather than taking block clear cuts. It also promotes managing for a woodland with a diverse range of tree species and ages.

Due to high prices of land with or zoned for structures, HAWC members decided to buy and move onto land and retroactively apply for planning permission. Over the ten years of the community, twice the Park Authority has denied residential planning permission, and both times HAWC has appealed the decision. Both appeals were granted largely on the grounds that HAWC provided an important and engaging experiment on sustainable lifestyles and woodland management. The most recent appeal was granted in the spring of 2009 and gave HAWC temporary planning permission to inhabit the site; however, this permission is due to expire after five years.

The Park Authority is keen to manage the woodland in a way that encourages increased biodiversity and species habitat, but generally the Authority is not convinced that the HAWC experiment is working. A number of forestry experts from the Authority have declared that HAWC has not done enough to promote the changeover (i.e. felling conifers and planting and coppicing important broadleaf species).

*Note: The name of the community has been changed.

Proposed Project

The ongoing debate between the members of HAWC and the Park Authority over the best methods of forest governance begs further investigation. On the one hand, HAWC recognises that they have not reached their woodland management potential largely due to the fact that it took a number of years for residents to settle, build houses, start gardens, and form routines. However, they likely have had an impact on the local forest ecosystem.

I would like to use remote sensing to help identify the extent and type of impact. Looking at land cover change over the past ten years via remote sensing will provide a dimension to the discussion that, to date, has been guided by the respective ‘expertise’ of the community members and Park Authority. I hope to be able to identify change in land cover trending away from conifer monocultures in favour of biodiverse broadleaf via selective harvesting methods. This would give the community support in future bids for planning permission. If no change in land cover can be detected, the project at least will provide a baseline from which HAWC can work in order to identify future progress.

Possible Methods and Data Sources

At the simplest level, I could use USGS Landsat images of the woodland area from the past ten years in a model using NDVI to look for land cover changes. However, there are a number of different iterations that may strengthen the analysis.

1) Alternative Land Cover Change Models: A paper by Wilson and Sader (2002) describes the use of Normalized Difference Moisture Index (NDMI) to identify land cover change due to partial harvest (as opposed to clear-cutting) management of temperate forests. Traditional NDVI analysis of partial harvest areas tends to have low accuracy. Wilson and Sader found significantly higher accuracies when using the NDMI method. This methodology may be highly applicable to the HAWC site where the focus has been on partial harvest techniques.

2) Comparison Areas: It is likely that there will be some variation among images from different years due to various radiometric effects. Because I likely will not have time in the scope of this project to do many corrections. To help compensate, I could use ‘control sites.’ If they exist, I would identify similar conifer plantation sites in the area that have not been managed to promote land cover change. Additionally analysing land cover change at a control site would provide a base line ‘null hypothesis’ situation.

Pros and Cons

The pros of this project are: 1) it potentially uses an innovative land cover change detection technique, 2) it is grounded in a relevant and extendible case study, 3) it has numerous possible levels of analysis (i.e. backup options), and 4) it has potential to actually benefit a group of people.

The cons are: 1) if data for extended analyses (NDMI and control sites) is not available, the project is just a case study replica of the lab from Week 5, 2) I am less comfortable starting projects in IDRISI than I am in ArcMap, and 3) I will be away from Oxford much of the break when I would be working on the project and I will not be able to use IDRISI in this time.

IIb. Actual Geographies of the Radical Rural: The Spatial Story of England’s Intentional Communities

III. Supervised Classification Lab

Supervised classification uses a researcher’s knowledge of an area to create ‘training sites.’ For example, if I had a map of Oxford, I would be able to draw a polygon around different features and classify them as buildings, fields, water, etc. Then, I could use the remote sensing IDRISI software to create a spectral signature (i.e. the specific colour reflectance values) of each of those land cover classes and have it apply corresponding classification to individual pixels with the same (or similar) spectral signature. In this way, even if I only am familiar with east Oxford, I could estimate land cover in north Oxford if it has similar properties to east Oxford. Here are the images and questions from the Supervised Classification Lab.

In the above image, I used what I knew about the site to draw polygons around areas I assumed were representative of each land cover type. I then had to make a signature for the average reflectance of the pixels in each polygon (Figure 4-3-2-1). In my first run, two of the land cover polygons did not encompass enough pixels to be significant (shown in red). Note, if you click on the image you get an up close look that is easier to read. I modified my polygons, and managed to get a significant polygon area for ‘Shallow Water.’ However, there was not enough area I knew to be agriculture to get a polygon large enough to significantly represent agriculture land cover.

I could then compare the signature of each specified land cover class. Each image is comprised of 7 bands of light, shown in Table 1. When each is plotted together, it can help identify which band will help to best differentiate land cover. For example, Figure 4-3-2-2 shows the plotted signatures of each identified land cover class (image on left). On bands where the lines overlap, the signatures are not easily distinguishable. However, in bands like h87tm4 (or near infrared)  the signatures do not overlap, indicating that this band would best differentiate the different vegetative covers.This function can also help identify characteristics about individual land cover classes. For example, the right-most image of Figure 4-3-2-2 shows a comparison of urban and coniferous forest land covers with the maximum and minimum values included in each polygon indicated. The urban land cover has more variation in reflectance values than the coniferous forest, probably because urban areas have a wide range of surfaces and substances (i.e. tarmac, pavement, lawn, buildings, etc).

Another way to compare and evaluate signatures is using a scatter plot that plots individual pixels based on two bands. Figure 4-3-2-3 shows a scatter plot of pixels based on the red (y axis) and near infrared (x axis) values of each. For example, a pixel with low reflectance of both red and near infrared light would register in the lower left-hand corner. Water falls in this category because water absorbs most light. The circles indicate which pixels belong to each category. Ideally, each pixel would fall only within one land category type. In this example, the signature circles do have some overlap points. Although I am sure I had some error in defining each land cover signature, some of the overlaps do generally make sense, i.e. shallow water and deep water have similar reflectance values.

Now that I have identified a signature for each type of land class, I want to use those signatures to identify the other pixels (outside my specified polygons) as a specific type of land cover. There are several models that use different functions to classify unknown pixels. I will not go into the technicalities of their differences and pros and cons. The images resulting from four different models and answers to accompanying lab questions are shown below.

Question: Compare results of using raw and normalised MINDIST functions. How would you describe the effect of standardizing the distances with the minimum distance to means classifier?

Answer: Generally, the normalised MINDIST classification had decreased areas of deciduous forests in favour of larger amounts of coniferous forests and urban areas.

Summarized actual answer: There will probably be greatest difference in urban class because of the heterogeneous nature of the urban land cover. Because the standard deviation will be high for the urban cover, it will classify pixels that are actually of another category to urban.

Question: How much did using standard deviations instead of minimum and maximum values affect the parallelpiped classification?

Answer: Effects of minmax piped versus z-score piped: Using the z-scores greatly decreased the number of pixels within the definable range. This leaves a large portion of the image’s land cover (namely, it seems, the coniferous and deciduous forest areas) unclassified.  

Summarized actual answer: z-score values are likely smaller than the min/max values, so some pixels will remain unclassified. However, change can also be attributed to the list order of land classes. If a pixel could be classified in multiple categories, it will be automatically assigned to the category higher on the list.

Going back over the ‘answers’ I can see that there are almost certainly places where coniferous forests decreased with the z-score answer due to the position of coniferous forests in the signature list.

Question: Of the above models, which classification is best?

Answer: It appears to me that the Fisher model does a very good analysis and seems to have the fewest ‘hanging pixels’ whose classification seems incongruous to the image (i.e. agriculture in the middle of a coniferous forest which is possible but unlikely). However, the ‘best’ classification probably depends on what an individual analysis requires. For example, if only rough, fast estimates are needed, the max/min parallelepiped classification may be best suited.

Summarized actual answer: Trick question! While some of the classifications are obviously erroneous, one can only identify the ‘best’ classification method by doing an accuracy assessment based on additional trips to the field.

I can see how my answer is inadequate based on the answer given in the guide. My response was based on an assumption that I knew more about the landscape than I actually do. For example, I liked classifiers that created fewer areas with single-pixel land cover types. However, these might actually be key areas for classification (i.e. looking for urbanised areas within forested regions).

IV. Unsupervised Classification Lab

Unsupervised classification requires no knowledge of a site to assign pixels into land cover clusters based on similarity of spectral signature. However, the produced image only shows clusters of like pixels — it does not identify what that cover is. In this sense, even unsupervised classification requires some degree of familiarity with the site. I initially used a broad clustering method to create classification groups. This produced six clusters (Figure 4-5-1). Given my knowledge from the supervised classification lab, I think these six clusters roughly represent:

1)      Deciduous forest

2)      ?

3)      Urban

4)      Water

5)      Agriculture

6)      Coniferous forest

Note, I think there is some other type or subtype of deciduous forest cover that is represented by cluster 2. Also note, this analysis does not distinguish between shallow and deep water.

I then used a finer clustering technique that distinguishes less significant clusters (Figure 4-5-2-1. This technique produced 35 clusters. Of them, [deep] water is the most easily identifiable, probably because it has the most different light signature, i.e. water absorbs all light. Other than possibly urban, the other land covers are different types of vegetation, presumably each species with its own signature.

While the first classification might have been too broad, this one is too fine. I used a histogram of pixel classification frequency to identify natural breaks in significant cluster sizes. I saw natural breaks at 5, 7, and 8 clusters; however, the lab identified natural breaks at 6, 10, or 15 clusters (Figure 4-5-2-2). Based on the lab’s suggestion, I reran the fine cluster analysis, limiting output to ten categories (Figure 4-5-2-3).

Although this limitation seems more accurate, I want to redefine land cover to more closely relate to the covers in the supervised classifications lab. To this end, I identified each cluster and recombined them into six new cluster types (Table 2) and reran the cluster analysis with the newly-assigned land cover classes (Figure 4-5-2-4).

*Note: you can click on the images to see the larger version in a separate window.

*For a document version of this report, click here.

CONTENTS (click each to be directed to that section)

I. Reading Summaries

II. Air Quality Lab

III. Burrowing Owls Lab

IV. Gap Analysis Lab

I. Reading Summaries

II. Air Quality Lab

To help assess needs and redundancies in the air quality monitoring system of EPA Region 9, I created a map summarising monitor data in California, Arizona, and Nevada. I decided to use carbon monoxide (CO) as an example pollutant. I first cleaned up the CO data to exclude measurements from Hawaii and low outliers (i.e. monitors reporting zero CO concentration). I then presented data points on a graduated colour scale (Fig. 1). I prefer this to a graduated symbol scale because it allows the viewer to better see each point, and I believe it is easier to immediately distinguish between several colours than several shape sizes.

To help determine redundancy of monitors, I calculated Moran’s I and Geary’s C using CrimeStat. I anticipate that Geary’s C will better represent our concerns (i.e. are neighbouring monitors measuring the statistically the same air?) because it deals with local clusters. Moran’s I value is less than 2.00 at 0.42, indicating that there is no autocorrelation among the points. However, Geary’s C value is less than 1.00, indicating that there in fact is an autocorrelation among the points. These results indicate that overall across the region, the monitoring stations are not redundant. However, at a local scale there is a significant level of redundancy. This trend within Moran’s I and Geary’s C is not surprising because one would not expect a monitoring station in Los Angeles to be measuring the same sample as in Phoenix, while the multiple monitoring stations in the greater Los Angeles area may be measuring the same statistical sample.

I made several interpolation maps using Kriging (Fig. 2), nearest neighbour (Fig. 3), spline (Fig. 4), IDW (Fig. 5), and bounded IDW (Fig. 6) techniques. In this case, the nearest neighbour looks the ‘most realistic’ because it has soft borders that roughly correspond to urban areas. However, the Kriging model is likely more representative of actual air pollution levels because of its use of variograms to extrapolate data. In air pollution, the variogram would be important because the nearest neighbour to a point might be very different (i.e. if one was next to a large point-source pollution site). The map would be even more accurate if one were to create a model variogram accounting for factors like predominant wind direction/speed or geographical features. Lacking data to make a more accurate variogram, I used the air basins of California to estimate local ‘boundaries’ to air pollution. It makes sense that air pollution would settle in natural basin formations that would therefore act as natural air pollution barriers. However, the interpolation map is still very skewed because even a basin feature would not entirely contain the pollution, but ArcMap interprets the barrier as an impassable object.

III. Burrowing Owls Lab

1. Mission 1, if you choose to accept it, is to spy on NASA

I accept your challenge, sir. First, I would like to report the coordinate system specs on the aerial photo acquired from the NASA site. A coordinate system is a scheme through which data may be projected in space. The image projection is in the North American Datum (NAD) 1983 UTM Zone 10N. In this projection, angles and shapes are more or less accurate, but distances and areas are distorted.

2. Whooooo let the owls out? Hoooo hoooo!

I used three methods to help determine degree of clustering of owl burrows on the NASA site in question: 1) estimation based on visual groupings; 2) statistical estimation based on Nearest Neighbour analysis; and 3) statistical estimation based on the Ripley’s K statistic.

The first estimate, visual grouping, produced nine possible clusters of burrows (Fig. 1) ranging in radius from 61.35 m to 141.07 m. When the two outlier clusters (containing only two burrow sites each) were eliminated, the average cluster size was 207.52 m in radius (Table 1).

3. Baby got stats

The second two estimates used the Nearest Neighbour and Ripley’s K functions (below are basic “layman’s” descriptions of each). In general, it is better to find statistically significant clusters for several reasons. First, statistical significance gives weight to an argument in the way that ‘eyeballed approximation’ does not. Second, it may help to identify exactly which points would go in what clusters which would affect the size of the cluster. This has management implications (i.e. size of mowed management areas).

Nearest Neighbour analysis (based on nine runs at various border types and with one, two, and five neighbours) produced a mean value of 73.02 (Table 2). However, because the plot is rectangular, the trials done with a “rectangle” border might give the best estimate (65.32). This is much lower than the non-statistical visual estimates.

Ripley’s K function analysis was also done in all three border types (none, rectangle, circle) and with various numbers of iterations (1, 10, 100). Each iteration was taken using a different distance value, so the resulting data is a list of n (1, 10, 100) distance values and the resulting L(t) values. The range where the L(t) value is above the L(t) maximum line represents the distances at which there is significant clustering, and the highest L(t) value corresponds to the distance (i.e. radius) at which the clustering is the most significant. The distance at which burrowing owls had the highest clustering is 202.7 m with significant clustering between approximately 35-490 m (Fig. 2). This is based on a trial run through 100 iterations on the rectangular border setting. Running the function through different numbers of iterations did not change the maximum L(t) value or distance at max value, but it did change the L(t) max and min lines which in turn changed the range of significance.

*Nearest neighbour: This is a basic tool for analysing clustering. It assumes three possible states for a set of points: 1) random, 2) clustered, or 3) dispersed. The tool calculates the distance from the centre of each feature to the centre of its nearest neighbour feature. It then averages these distances for the sample and compares it to the average distance for a randomly distributed sample. If distances are lower than the random sample distance, the points would be closer together, indicating a clustered pattern. If distances are higher than random, it indicates that points are dispersed. For example, look at Wyoming. Let’s say you think our neighbourhood represents a good (random) distance from your neighbour, about 0.5 km. In town, houses are only about 20 metres from each other. Because this number is lower than 0.5 km, the houses in town are considered clustered (which makes sense). However, out in the ranch lands, houses might be 40 km apart, a distance greater than our own ideal 0.5 km, so houses here are considered to be dispersed.

*Ripley’s K: This tool analyses clustering by counting the number of neighbours of a feature at different distances. If at one distance from the feature there are more neighbours than there are at an average concentration throughout the sample, Ripley’s K denotes a cluster at that level. Using the same Wyoming example, start at our house and make a series of concentric rings around it. At each ring distance, count up the number of houses (neighbours) within that boundary. Up to about 10 km radius from the house, we have our random ideal house density. However, if we compare to a house in town, we see that there are more neighbours than the ideal density within a polygon of around 15 km radius. Thus, again, town represents a cluster but we know more about the cluster (relative location and size).

4. Management madness!

Ripley’s K analysis for the NASA site before conservation management mowing (March 2004) and after mowing (August 2004) shows how owl behaviour changed due to conservation management actions. Clustering in March was of low significance with a significant maximum L(t) value of 130.2 corresponding to significant clustering activity at 42.7 m (Fig. 3). However, post-mowing behaviour indicates a higher degree of clustering at a broader scale of (higher) distances with maximum clustering behaviour at 101.9 m (Fig. 4). This indicates that mowing may have allowed the owls to move into more space in a more definitive clustering pattern (possibly corresponding to mowed spaces?)

Some additional factors that may be taken into account in deciding management areas are spaces of vegetation and predator/prey distribution (Fig. 5). Most of the owls were sighted in vegetated areas, so conservation measures may be most successful if they focus on similarly vegetated areas. It is also important for habitat supporting owls’ prey is conserved and accessible to the owl population. This means that although the field in the northwest corner of the plot did not have any owl sightings, it should be minimally disturbed to protect the owls’ rodent food source. Additionally, managers may want to undertake additional protection measures for the owls in areas where predators frequent.

This analysis overall is strong in that it uses several methods of estimating clustering of burrowing owls under several environmental conditions (mowed/un-mowed vegetation). Looking at clustering (as opposed to absolute numbers) allows better conceptualisation of owl movement and proportional representation in mowed/un-mowed areas as opposed to giving snapshot numbers that are subject to change by overall population flux. The study is less strong in that there is room for GPS error (owl location could only be estimated from a distance), it cannot directly correlate management actions to owl behaviours, and it is based on only one year’s data in one location.

IV. Gap Analysis

This lab looks at the technique of gap analysis as a potential conservation tool for species in California. I chose to focus on several raptors, three hawk species and three species of falcon. Predicted  Swainson’s (threatened), Red-Tailed, and Red-Shouldered hawk habitat, based on land-cover types suitable for each species, can be seen in Figure 1. Predicted Kestrel, Peregrine (recently delisted), and Prairie falcon habitat range can be seen in Figure 2.

Gap analysis can help to identify biodiversity hotspots, particularly those that remain unprotected. This is an important step forward in conservation management because it promotes proactive conservation of important habitat as opposed to retroactively trying to protect species that are already threatened. It helps to focus conservation on biodiversity, which is a better measurement of overall ecosystem health than the promotion of a single species (i.e. “charismatic megafauna.”) Much of the weakness of gap analysis lies in the inherent imprecision of maps. For example, vegetation maps cannot capture important landscape features like micro-habitats, vegetation quality, actual (as opposed to estimated) animal distribution, or the gradual borders between vegetation types. All of these are important factors to consider when proposing conservation management strategies.

Because I limited my analysis to six bird species due to time constraints, I have limited ability to comment on the correspondence of areas of species richness, endemism, or critical habitat to threatened species. However, the analysis did demonstrate several key points. First, although the groups are similar (i.e. all hawks/falcons), there are still limited areas (particularly among the hawks) where all three species are expected to coexist, i.e. hawk/falcon biodiversity hotspots. Second, many of these hotspots are not yet protected (Fig. 3 & 4). In particular, the amount of land recognised as unprotected falcon hotspots outweighs protected falcon hotspots by a third (Table 1).

The third key point is that focusing only on species of concern, i.e. Swainsons hawks and Peregrine falcons, does not account for all of the land identified as hawk/falcon biodiversity hotspots (Fig. 5). While focusing on special or indicator species or land-cover types may protect portions of biodiversity hotspots, that style of conservation management is not likely to proactively protect entire spaces of ecological importance (biodiversity).

*Note: you can click on the images to see the larger version in a separate window.

*For a document version of this report, click here.

CONTENTS (click each to be directed to that section)

I. Reading Summaries

II. Introducing ArcMap Exercise

III. Wytham Woods Lab

I. Reading Summaries

Scott et al (1993) Gap Analysis: A geographic approach to protection of biological diversity

The paper by Scott et al presents an historical (16 years old) but in-depth overview of the concept of gap analysis via GIS. Gap analysis is a process of identifying conservation hotspots (i.e. sites with the potential to host many types of animals and plants) and using them to determine where there are existing gaps in conservation efforts. It uses GIS software to relate the vegetation of an area to the suitability of that area to host different animal species. This is done by first classifying the vegetation of a particular location using actual observed data or estimating cover from images. Second, the vegetation information is used to evaluate areas of suitable habitat for various animal species based on what is known about each species (i.e. what does a species need for food, shelter, etc.) This can help predict animal distribution and biodiversity. Then, the maps of distribution and biodiversity can be compared with maps of protected and unprotected areas to help identify areas for future conservation efforts.

Personally, this paper helped me to understand the differences between vector and raster data (both their origin and uses). I also learned how some gap analyses (i.e. wetland management and rare species) must be treated differently due to different abilities, parameters, and assumptions of the GIS software. Although there were some fuzzy points due to my lack of familiarity with GIS and remote sensing and the length of the paper (i.e. I didn’t read everything to the depth I normally would), I felt that it was clearly written and left me with few questions.

Short Essay Question: In what ways can gap analysis contribute to conservation efforts?

Short Answer: At a base level, gap analysis identifies potential key habitats for various species. This is important for several reasons. First, animals are mobile and so an image (i.e. a satellite image) cannot demonstrate on its own the range and habitat for an animal species. Second, it can show where conservation efforts may be needed to protect habitat before its reliant species become threatened or endangered. For example, it can show habitat suitable for species on the brink of listing (like the sage grouse in Wyoming) and identify which portions of that habitat have already been protected and which may still need further protection measures. This is an important step forward in conservation practice because past conservation measures have been retroactive, meaning conservation efforts have attempted to help a species once it is already in trouble, rather than proactively preventing the crisis by preserving habitats, particularly those mostly likely to sport high plant and animal biodiversity.

Hinton (1996) GIS and remote sensing integration for environmental application

This is a review paper discussing the importance and environmental applications of integrating GIS and remote sensing analysis systems. GIS systems tend to deal primarily with vector images, or discrete points, lines, and polygons. Remote sensing systems tend to deal primarily with raster images, or images that are externally ‘sensed’ via camera, radar, or satellite (to name a few). These images are constructed of many individual internally uniform pixels, each with its own set of properties. Traditionally, GIS and remote sensing images could only be combined via the conversion of data from raster to vector or vector to raster format. These conversions often cause serious errors. This paper reviews studies concerned with the challenges and applications of systems that facilitate integration of the two types of data. According to Hinton, the primary applications of integrated systems would be 1) using GIS to process remote sensing images, 2) using remote sensing images as GIS data sources, and 3) a reduction of errors caused by converting data formats (vector to raster and vice versa).

When I originally read this paper in week one of the course, it was most useful in helping me to understand the difference between raster and vector data and the different applications of each. I particularly found it interesting that while some of the concerns presented have been addressed in the past thirteen years, the systems are still far from full integration. It was interesting to look back over my notes from week one knowing what I know now. At the time, I did not understand the section about remote sensing image classification using GIS or vector data. Now I better understand the process. It seems as though current methods of image classification are more integrated than they were in 1996, and I can appreciate the importance of being able to use vector polygons to identify raster features (i.e. in the selection of training areas).

Short essay question: Describe several planning or conservation situations in which it is most appropriate to use 1) vector images, 2) raster images, and 3) integrated systems.

Short answer: Imagine a study looking at the effects of urban sprawl (based on building density) on pronghorn antelope habitat. Buildings may be best represented by vector points or polygons because they have discrete edges that may be blurred or hard to distinguish in raster format. However, suitable habitat for pronghorn will not tend to have discrete edges. Short grass prairie will be mixed to different degrees with sage prairie and interspersed with clusters of different tree varieties. These habitats may be better represented by raster images. However, in the identification of suitable habitat and to analyse the interaction of development (represented as vectors) and said habitat (represented as rasters), it would be most useful to have an integrated system that could utilise, combine, and analyse both data formats.

Fisher et al (2007) An Analysis of Spatial Clustering and Implications for Wildlife Management

This paper reviews a study done on a NASA test site in California concerning conservation management techniques for burrowing owls. Specifically, the study tested the ability of a certain statistical analysis, Ripley’s K function, to determine changes in clustering of nest sites in response to mowing, a typical conservation measure. This is the first step in determining if mowing is an effective conservation tool (i.e. answering the question, “Do more owls live and nest in areas that have been managed and mowed?”) Previous studies have used other analyses to determine clustering, but according to Fisher et al these analyses have been insufficient. One, GIS spatial analysis, looks for clustering at set, discrete distances. The other analysis measurers distance from any one point to its nearest neighbour, but it cannot assess clustering beyond this. Ripley’s K, however, is more versatile because it assesses clustering at all ranges among all points. The study found that owl nests were randomly distributed within the study areas in the year no mowing occurred and were clustered in the year when the area had been mowed. Cluster sizes also roughly corresponded to mowed area sizes. However, because this was not set up with a control or over an extended period of time, the study cannot statistically determine that mowing caused clustering, only that Ripley’s K can identify a change in cluster patterns.

I found this to be a very useful and interesting example on how GIS can be used to help conservation management techniques or policies. I found it particularly interesting to think about the ways that different statistical measures can have different consequences and outcomes. At the time that I read the paper, I did not understand the Ripley’s K function very well. I felt that while I understood the implications and inputs of the model, I did not understand the means for deciding significance. I think I understand this much better now that I have done the Burrowing Owl lab!

Short essay question: Describe how GIS can affect conservation management practice.

Short answer (and I mean short! You could write a book on this topic!): Best management practices for conservation can be determined by any number of factors from politics to stakeholder needs to scientific (including GIS) studies. GIS analyses in particular can contribute a unique spatial dimension to these studies. Many ecosystem studies focus metrics such as biodiversity, species presence, species abundance, or fecundity (to name a few) in response to various environmental pressures, including management techniques. However, these metrics cannot capture certain spatial patterning like clustering. One case study utilising GIS to better understand the effects of conservation efforts on a species is discussed by Fisher et al (2007). In this study, Fisher et al use the Ripley’s K function to analyse clustering of burrowing owl nest sites in response to mowing, a best conservation management practice. While the study could not statistically attribute clustering to the practice, they did find correlation of increased clustering in post-mowing environments (compared to the same area pre-mowing) with cluster sizes roughly correlating to the sizes of areas mowed. The findings of this study reinforce the current best management practices for the owl; however, if no correlation or negative correlation had been found, the study may have contributed to changing the policy for future management areas.

II. Introducing ArcMap Exercise

A map of landmarks and American Indian reservations in Hot Spring and Freemont counties in central Wyoming.

III. Wytham Woods Lab

I. Coordinate systems and projections

The coordinate system used for the Cloud Free Earth image is GCS_WBS_1984. It is important to know the coordinate system/projection for layers in GIS because different coordinate systems can cause data to be represented differently, i.e. image distortions or measurements changes. For example, in the GCS_WBS_1984 coordinate system, the study site boundaries measured approximately 165 m x 53 m which is acceptable based on the reported GPS error. However, when the image is displayed with projected coordinate systems as opposed to geographic projection coordinate systems (i.e. World Craster Parabolic, World Cylindrical Equal Area, North America Equidistant Conic), the image appeared in the wrong location (i.e. Africa or the USA) and had measurements at a fraction of the actual site (i.e. 0.002 m x 0.000534 m). The different projections change distance measurements because each ‘stretches’ or ‘wrinkles’ the surface of the spherical earth to show it in a two-dimensional representation. If these stretches or wrinkles appear within the measured space when switching among different projections, the measurements of that space will change.

II. Forest data

The study site in Wytham wood, based on colour-coordinated groupings, appears to have been marked out by five groups (1, 2, 4, 7, 8 ) marking out the north and south borders of the rectangular plot. Each group measured half of the border with groups 4 and 7 along the same section but with slightly different measurement points.
The Wytham remote sensing image initially appeared near Africa because it had not had a projection or coordinate system set for it. The image after repositioning using coordinate georeferencing can be seen in Fig. 1. The overlaid GoogleEarth image of Wytham wood was repositioned using five ground control georeference points (Fig. 2).
The forest data (tree species, height, diameter, etc.) helps illuminate a number of characteristics of the wood. When overlaid on the GoogleEarth image, one can see gaps in tree points that correlate with clearings on the image, as represented by darker spaces among the green foliage (Fig. 3). Many species appear to have a generally scattered distribution. However, some (i.e. field maple and blackthorn) tend to cluster together. Blackthorn in particular seems to cluster together and near clearings (i.e. on the border with the field). At a glance, the highest biodiversity appears to be in the northeast corner, particularly along the edge of the forest. However, the diversity indices indicate that while tree abundance is higher in the northeast corner and along the forest edge, biodiversity is higher further away from the edge.
	Sizes of trees are diverse and appear to have moderate spatial variation. Trees with smaller diameters (Fig. 4), heights (Fig. 5), and biomass (Fig. 6) seem to be predominant next to the northeast edge and within/along clearings. Otherwise, trees of moderate to large size are fairly evenly distributed. There seems to be a general correlation of diameter, height, and biomass. The highest correlation seems to be among trees of small size. Although trees with ‘high’ and ‘medium’ diameter do not always correspond exactly with trees of ‘high’ and ‘medium ‘height and biomass exactly, this is probably at least partially due to my random categorical selection of low/med/high measurements (Fig. 7, 8 ).