Position on the anthropocline predicts the type of EH-BD relationship

Applications of computer to measure complex ecological properties and patterns – moving beyond single-feature metrics.

In Chapter 3, l investigate and develop new automated tools for rapidly quantifying ecosystem features and visual patterns. The data derived from such metrics, or in sorne cases we may term them « ecological indicators« , may en able us to fur~ ther study andunderstand biodiversity-EF relationships and, in turn, improve the methods we use to manage and conserve ecosystem services. The methods offer finer-scale spatial and temporal resolutions over current datasets, while achieving high rates of throughput, cost and time efficiently. This research objective is initiated in Chapter 2 in which l apply time-lapse digital photography and derive unsupervised, single-feature metrics from them (e.g. , greenness indices) to quantify the seasonal growth patterns of forest stands. In Chapter 3 l move beyond these single-feature-based image metrics and look to the rapidly advancing field of computer vision to apply more complex, supervised learning models. These methods use multiple image features, either computed as bags-of-features or generated by artificial neural networks, to first segment environmental real-world features represented in an image (e.g., trees) and, second, derive and quantify useful measures of those features (e .g. , pe.rcent tree cover).

As the accessibility and capabilities of digital imaging continue to rapidly grow, researchers have begun to explore the merits of digital photography in various ecological applications (Graham et al. , 2010; Granados et al., 2013; Mellin et al. , 2012; Proulx et al. , 2014; Crimmins et Crimmins, 2008; Olea et Mateo-Tomas, 2013; Rousselet et al. , 2013). A key feature of digital photographs is that they can be mathematically manipulated in numerous ways to detect and enhance patterns and classify objects. What’s more, analytical and mathematical advances have been complimented with advances in computing power and data storage. Thus, if images and image-based metrics can be shown to represent ecosystem properties that have traditionally been measured and collected with difficulty (i.e. , involving large amounts of resources, man power and time) , their value becomes obvious. What’s more, mathematical derivations of environmental properties, re- . presented in images, may also be able to capture higher-Ievel ecosystem features that are otherwise very difficult to quantify and describe (e.g. , heterogeneity of light, structural diversity or heterogeneity, leaf angle). Computational power and data technologies also enables researchers to collect, aggregate and analyze high volumes of image data captured at very high levels of temporal and spatial resolution; i.e. , high-throughput at fine-grain resolutions with the potential to be applied at large extents.

Beyond single-feature image metrics, computer vision scientists are teaching computers to see and under stand the world at astounding rates of success. However few disciplines outside of the strict artificial intelligence fields (e.g. , robotics, medical devices, driverless cars, web analytics) have utilized these advancements. Yet, if computer vision algorithms can learn to detect features of an environmental scene, it stands that those algorithms can be used to objectively quantify realworld features and their spatial distribution within a landscape for a multitude of applications. Higher-Ievel ecosystem properties (e.g., forest stand complexity) have traditionally been quantified through many manual measurements and labour intensive multi-step analyses. A powerful advantage of machine learning methods to qu ant if y environmental patterns is the ability to identify complex and non-linear patterns that may not be observable otherwise or only intuitively so. The specifie objective of the third research chapter will be to apply computer vision tools to begin to develop automated assessment tools to measure, map and track environmental features of ecological importance. Specifically, l will apply computer vision algorithms that have become well-accepted within their own field (Hoiem et al., 2005; Naik et al., 2014a) to map and measure trees. To do so in an urban landscape context, l will also utilize the open-source image data of city streetscapes that is now abundant – Google StreetView Images (Google Inc. , 2014)

Revisi ting environmental heterogenei ty

Perhaps the first pressing problem when it cornes to understanding EH-BD relationships is how to define and measure heterogeneity. Traditionally, EH has been measured on an unbounded gradient ranging from low to high spatial or temporal variation, with the observed values ranging somewhere in between these extremes (Fig.l.l). For instance, most measures of environmental variation for a given area (e.g., coefficients of variation, standard deviation, landscape texture, edginess, interspersion metrics) adhere to this definition. However, a limitation arises in that the results of any such attempt to quantify the EH-BD relationship lack a reference point on the heterogeneity gradient, thus preventing cross-study comparisons. Since the EH gradient not only exists within ecosystems, but spans across all ecosystems and levels of the human footprint , the second pressing problem is how to organize the EH gradient between ecosystems. Proulx and Parrott (Proulx et Parrott, 2009; Parrott, 2010) proposed that natural ecosystems (i.e. , ecosystems subject to no, or little, human modification) should be considered neither uniformly organized nor completely disordered, but rather as « complex » systems.

Accordingly, natural ecosystems should fit between the two heterogeneity extremes, at sorne intermediate level of EH (Fig.l.l). Indeed, environmental conditions within natural ecosystems typically reveal sorne intermediate level of spatial or temporal patterning (i.e. , are neither uniform or disordered) at one or several scales of observation (Legendre et Legendre, 1998). For example, in a recent study Seiferling et al. (Seiferling et al. , 2012) quantified the spatio-temporal landscape heteroge29 neity of vegetation cover both inside and outside 114 large protected areas around the world. While the pattern of vegetation cover was consistently more heterogeneous outside protected areas, the results revealed that the heterogeneity inside their borders averaged around c.a. 0.5 on a scale bounded between 0 (complete uniformity) and 1 (complete disorder). In other words, the patterns of natural vegetation cover within protected natural areas presented a great deal of EH, well above uniformity, yet well below complete disorder.

A new conceptual framework for the EH-BD relationship

In acknowledging that natural ecosystems retain sorne intermediate levels of EH, we may consider what happens to such an ecosystem if the amount of human modification is increased. Initially, novel environmental conditions are created at the expense of pre-existing natural ones (i.e. , a human expansion phase ; Fig.1.1) . The conversion of natural lands to small farms, low levels of urbanization, road development, or the dissemination of diffuse acoustic and chemical stressors, along with the alteration of food-web dynamics through resource harvesting would all constitute examples of this modification. The net result of this human expansion will, generally, be an increase of EH. During this phase, species diversity in several taxonomic groups may decline because species with larger home ranges or specialized resource requirements may be extirpated and, therefore, negative EHBD relationships dominate. Even in situations where multi-species coexistence in patchy habitats appears to persist, species may exhibit extinction debts wherein local extinction is just a matter of time (Tilman et May, 1994; Vellend et al. , 2006).

As the human footprint further increases from early phases of semi-natural states, natural habitats dwindle in size and frequency, pushing the ecosystem into a phase of human intensification (i.e. , human modification of ecosystems from semi30 natural to highly-modified states; Fig.1.l). Farms become farmlands and towns are replaced by high levels of urbanization, leaving behind only remnants of the original ecosystems. During this phase, biodiversity may further decline because only a few species are adapted to live in such homogenized habitats or are able to sustain the environmental stressors for long periods. Additionally, coexistence among competitor species may no longer be possible given the lack of EH, driving further losses of biodiversity through competitive exclusion (Amarasekare, 2003). atural ecosystems can also be converted directly to intensively modified states by pro cesses like clear-cutting forest for agriculture or urban development (i.e. , rapid homogenization ; Fig.1.l). State shift theory, for example, would categorize such rapid homogenization as a « sledgehammer » effect (Barnosky et al. , 2012). Nonetheless, as the pattern of resources, biota and structures in ecosystems becomes more uniform, eventually biodiversity collapses. Examples of such highly-modified ecosystems include intensive agricultural lands, or very densely populated areas with extensive paved and building cover. The EH-BD relationships in such homogenized ecosystems are now positive since an addition of EH increases habitat or resource opportunities, which should promote niche partitioning and species coexistence (Fig.l.l).