← projects

Summary:

An effort to assess the rela­tion­ship and poten­tial syn­er­gies of individual-​​based ecology and remote sensing, and to identify some of the specific chal­lenges of gath­ering remote-​​sensing data to develop individual-​​based eco­log­ical theories.

Context:

This was orig­i­nally written for an Introduction to Complex Systems course I took at U. Michigan’s Center for the Study of Complex System. It was written pri­marily between the hours of 8pm and 5am the night before it was due, and con­se­quently is on the ranty and ragged side, but it nonethe­less contains ideas I find inter­esting and poten­tially significant.

The chal­lenges outlined in this paper moti­vated my thesis research project.

Other Versions

In addition to the paper below, I deliv­ered a short pre­sen­ta­tion covering a subset of the content:

A .pdf version of the paper is avail­able here.

Content:

1. Individual Based Ecology
1. IBMs and IBE
2. Mechanism in modeling
3. Novelty and mechanism.
4. Prediction and generalizability.
5. Focus on the indi­vidual.
2. Remote Sensing and Ecology
1. Data chal­lenges in ecology.
2. The poten­tial for remote sensing.
3. Remote Sensing and Individual-​​Based Ecology
3. Sensors, Systems, and Scales
1. Choosing a sensor and a system.
2. Challenges: spatial resolution.
3. Challenges: temporal resolution.
4. Challenges: temporal scope.
5. Challenges: resolv­able system characteristics.
6. Challenges: cost and access.
7. Challenges: spatial scope.
4. Case Studies
1. Case study one: treeline development.
2. Case study two: inter­tidal dynamics.
5. References

Individual Based Ecology

IBMs and IBE.

In their book Individual-​​based Modeling in Ecology (2005) veteran individual-​​based modelers Grimm and Railsback propose a new name for a newish dis­ci­pline. “Individual-​​based ecology” (IBE) is to be a com­ple­men­tary arm of the eco­log­ical effort, focusing explic­itly on how individual-​​level dynamics produce, in aggre­gate, ecosystem-​​level dynamics. This new dis­ci­pline would be an effort to for­malize and extend method­olo­gies and theories which have their ad hoc roots in the devel­op­ment of individual-​​based models (IBMs). IBMs, in turn, are a class of models defined by their explicit encoding of the behav­iors of indi­vid­uals. They commonly utilize sim­u­la­tion in addition to, or in place of, math­e­mat­ical analysis. In both their focus on indi­vid­uals over systems and in sim­u­la­tion instead of purely ana­lyt­ical inves­ti­ga­tion, they differ from many decades of tra­di­tional eco­log­ical modeling work.

Ecological modeling dates back to the “golden age” of the­o­ret­ical ecology, when Lotka and Volterra had their coin­ci­dent epipha­nies about the poten­tial for math to guide insight into natural systems (Lotka 1932, Volterra 1927). Since then sub­stan­tial effort has been spent on attempting to explain or predict pop­u­la­tions and com­mu­ni­ties of organ­isms in natural settings. These models have commonly ref­er­enced the indi­vid­uals organ­isms in those settings only to the extent that they might use whole numbers to describe their quantity (in some cases in fact, the specific math­e­matic struc­ture of eco­log­ical equa­tions allows for frac­tional or even negative indi­vid­uals to exist). While such systems-​​based modeling has con­tributed to many exciting con­cep­tual triumphs, ecosys­tems remain only occa­sion­ally explain­able, and largely unpre­dictable. The scale of human-​​induced change in the natural world has made such expla­na­tion and pre­dic­tion increas­ingly desir­able. Both are nec­es­sary con­trib­u­tors to rational policy devel­op­ment, and con­ser­va­tion policy will only become more critical as mod­i­fi­ca­tion of natural systems con­tinues its seem­ingly inex­orable increase in speed and scope. Individual-​​based ecology rep­re­sents a fresh and provoca­tive addition to eco­log­ical thinking, and a poten­tial con­trib­utor to eco­log­ical expla­na­tion and pre­dic­tion. IBMs have grown in parallel with the devel­op­ment of complex-​​systems theory, a wildly cross-​​disciplinary effort to study systems as the con­se­quence of their inter­ac­tive parts. IBMs have both drawn from and con­tributed to the field of complex systems, and com­plexity and complex adaptive system (cas) theory rep­re­sents a deep and deep­ening well of ideas, moti­va­tion, and synergy for IBE. (Hartvigsen et al 1998, Levin 1999, Hirsch and Gordon 2001, Green et al 2005).

Mechanism in modeling.

Any model of a real system can have several criteria, and often success in one of these criteria can only be had by sac­ri­ficing in another. Prevalent in these trade-​​offs is the tension between pre­dic­tion and expla­na­tion (Brown et al 2005, 2006). A model may sta­tis­ti­cally predict the outcome of a system without making any ref­er­ence to the real com­po­nents of that system. Alternatively, a model may faith­fully reflect each of the relevant entities which build a system in real life, but may repro­duce the behavior of the system only qual­i­ta­tively, without pre­ci­sion. In practice, the nuances of model building compel the modeler to focus on one or the other of these goals. The advan­tage to pre­dic­tive modeling is obvious: pre­dic­tion is often our imme­diate goal in studying the world, so that we may develop appro­priate plans to attain or respond to a probable scenario. On the other hand, the advan­tage to an explana­tory model is that, by rep­re­senting the system as we truly believe it to be, we can advance and test our theories about the actual mech­a­nisms of that system. Any improve­ment in the pre­dic­tive capacity of an explana­tory model rep­re­sents a possible increase in our knowl­edge of how the real world func­tions, whereas increasing the pre­ci­sion of a black box pre­dictor only tells us what is going to happen better. If we are driven only by serendip­i­tous sta­tis­tics, then our capacity to improve our models is limited by our capacity to gather more training data and to improve our sta­tis­tical capacity to cor­re­late events. While we may rely to a degree on both of these devel­op­ments, a more fecund hope is that the human capacity for creative pattern recog­ni­tion will produce radical improve­ments in mech­a­nistic under­standing of a system. Predictive models are in any case typ­i­cally built upon explana­tory under­standing. So while pre­dic­tive models rep­re­sent poten­tially lengthy branches off the tree of knowl­edge, explana­tory models extend the trunk itself and provide the pos­si­bility of higher level pre­dic­tors to come.

One of the advan­tages of Individual-​​based models, and individual-​​based ecology more broadly, is that they are par­tic­u­larly suited to mech­a­nistic under­standing. Mechanistic fidelity comes from a tight­ening of the rela­tion­ship between the entities of a model (either a formal or purely con­cep­tual model) and the respec­tive entities in nature. By contrast, we can also imagine a purely sta­tis­tical and pre­dic­tive rela­tion­ship between models and nature. In this case, the content of a model is irrel­e­vant. It might be an equa­tional black box or it might have nar­ra­tive com­po­nents, yet be entirely dis­sim­ilar to the com­po­nents of the natural system the model is meant to predict. As long as the model predicts what the natural system is going to do, its internal struc­ture is moot. Product over process. A mech­a­nistic model iden­ti­fies what the most relevant entities are in nature and tries to describe and/​or mimic those entities as part of its structure.

What the most “relevant” parts of a natural system are will inevitably be guided by our specific and sub­jec­tive goals. Specific entities are (we are likely to claim) causally linked to specific dynamics. The success of individual-​​based approaches to date, and the com­pelling insights offered by com­plexity research in general, suggests that many of the specific dynamics we might be inter­ested in are most closely causally linked to the entities we think of as ‘indi­vid­uals’ or ‘agents’. We shouldn’t think that we are trying to obtain absolute gran­u­larity in our mech­a­nism. As Volker and Railsback point out, “every­thing that organ­isms are and do emerges from inter­ac­tions among their genes, their neurons, and their envi­ron­ment” (p. 118), and hence if complete mech­a­nistic fidelity were our bench­mark, then we would have intractably complex models. But as our models become increas­ingly individual-​​oriented, they are likely becoming more mech­a­nistic (Grimm 1999, DeAngelis and Gross 1992). Some drivers in ecosys­tems truly are system level: energy and nutrient flows tied to climate, soils and hydrology will carry on regard­less of the behavior of one organism or another. Those inter­ac­tions can poten­tially be encap­su­lated in a individually-​​mechanistic model. Many biotic dynamics are the result of individual-​​level inter­ac­tions, and those can only be captured by individually-​​mechanistic models.

Novelty and mechanism.

I suspect that one of the advan­tages of mech­a­nis­ti­cally real­istic models is that they have a robust­ness in the face of novelty. Statistical—predictive—models are composed of infer­en­tial rela­tion­ships between the data on hand and the sta­tis­tics we use to tie them together. As a product of numbers, they are espe­cially vul­ner­able to over­fit­ting, and can only be trusted to function within the range of con­di­tions for which training data was avail­able to para­me­terize them. Mechanistic–explanatory–models have a special capacity. If they are good mech­a­nistic models and reflect the true entities in the world, then they will behave like those entities, even if they are pre­sented with per­tur­ba­tions beyond the ranges of the data used to con­struct them. Hence, they may predict the outcomes of real-​​world sce­narios which we have not yet expe­ri­enced. Given that human impacts on the ecosystem are mod­i­fying system drivers beyond the ranges over which they have pre­vi­ously varied, having models which may predict real behavior outside of the ranges of observed data may be a very useful thing. Most strik­ingly, global climate change is likely to produce sets of patterns in ecosys­tems at many scales, patterns which may often be new to human expe­ri­ence (e.g. Niemann and Visintini 2005). Having models robust to such novelty should probably then be a goal for ecology (e.g. Maytin et al 1995). Thus mech­a­nism should be a goal, and thus individual-​​based ecology seems yet more relevant.

Prediction and Generalizability.

In his A Critique for Ecology, Peters (1991) blames his crit­i­cism that ecology is not gaining pre­dic­tive power on too great of a focus on mech­a­nism. He takes the pos­i­tivist view that assigning causes to effects is a vain pursuit, at least in the face of the massive number of poten­tially hidden actors in eco­log­ical systems, and suggests that we must focus our efforts on cor­re­la­tion over cau­sa­tion. Ecology has long strug­gled to produce rela­tion­ships which are both pre­cisely pre­dic­tive and also gen­er­al­iz­able from spatial location to spatial location, or across scales or among organ­isms. A science which requires a spe­cialist for every species and region in order to offer pre­dic­tion is not a very useful science. According to Peters, a research focus on sta­tis­tical infer­ence will lead to greater sta­tis­tical infer­ence, and sta­tis­tical infer­ence is what ecology is lacking. Explanation he likens to a glo­ri­fied natural history approach.

“In practice, research into causal con­nec­tions is not the same as the search for pre­dic­tive power and causal research programs assume some very dif­ferent char­ac­ter­is­tics. The crux of this dif­fer­ence is that causalist research seeks first to place the effect in an explana­tory model which will also provide pre­dic­tions. As a result, causalist research seeks first to build a causal web, rel­e­gating an eval­u­a­tion of pre­dic­tive power to such time as the web is described. Instrumentalist science strives for pre­dic­tive power from the first” (Peters 1991 p. 133)

Given his focus on pre­dic­tive power, Peters would probably be uncom­fort­able with individual-​​based approaches to ecology. IBE, as with other complexity-​​based dis­ci­plines, draws our atten­tion to the ubiquity of non-​​linear dynamics, and the con­se­quent path-​​dependant sen­si­tivity of those systems (Solé et al 1999, Levin 2002). Consequently we appear less, rather than more, likely to accom­plish pre­dic­tion. In par­tic­ular, our con­clu­sions seem to become less gen­er­al­iz­able (Batty and Torrens 2001). Any given mech­a­nism may yield sig­nif­i­cantly dif­ferent results from one place to the next, or even from one moment to the next. This would seem to be an anathema to Peters, an almost dialectic divorce from his intended direc­tion for eco­log­ical science.

While I agree with Peters that ecology must improve, or at least sophis­ti­cate (Brown et al 2005), its pre­dic­tive ability in order to be relevant to con­ser­va­tion planning, I am swayed by the con­trasting argu­ments of Volker and Railsback. They suggest that the unal­loyed desire for gen­er­al­iz­ability which has devel­oped in sympathy with Peters has cramped ecology into a posture inap­pro­priate for its task.

Ecological science in the 20^th century was marked by tension between tra­di­tional, descrip­tive, ‘natural history’ approaches and those who entered the field because they believed that they could pro­duc­tively apply their expe­ri­ences in the hard sciences of physics and math. In what strikes me as both a response to Peters specif­i­cally and a rallying cry for the unity of the two eco­log­ical schools, Grimm and Railsback reflect that

“Intuitively, we think of “general” theory as being inde­pen­dent of specific contexts. Theory in physics is very general in this way, but physics deals with matter and forces, which are indeed inde­pen­dent of history and context. Seeking the same sort of gen­er­ality in ecology has not proved to be very productive….Organisms are not atoms, and the “forces” of ecology are not fun­da­mental prop­er­ties of matter and space but emerge from inter­ac­tion among indi­vid­uals and their envi­ron­ment. In ecology, there­fore, useful theories are likely to be context-​​specific and the search for gen­er­ality must include the search for the limits of this gen­er­ality”. (2005 p. 56 –57)

Coming to grips with the limits of pre­dic­tion sug­gested by complex dynamics remains a major barrier to individual-​​based ecology, but one which may be revealed rather than created by it. Already we are devel­oping strong answers to ques­tions about generalizability.

Focus on the indi­vidual.

By putting “indi­vidual” in the proposed name of the dis­ci­pline, Volker and Railsback are going further than saying that it will be based on the method­olo­gies of com­plexity, or of cas. Many processes from other dis­ci­plines, which are iden­ti­fied as complex adaptive systems because of their focus on the emer­gence of system prop­er­ties from the inter­ac­tions of agents, do not nec­es­sarily have as their ‘agent’ an indi­vidual. In some cases this is because they operate at sub-​​individual scales, i.e. in physics or neu­rology, and in some instances, systems which include indi­vid­uals are modeled using some aggre­gate level of orga­ni­za­tion as the agent. For instance, In human ecology, the ‘house­hold’ is commonly the unit (e.g. Clarke et al 1997) , and in some existing complexity-​​based eco­log­ical models, ‘grid cells’ full of indi­vid­uals (Moloney and Levin, 1996), metapop­u­la­tions (Guichard 2005), or whole species (Nuetel et al 2002) may make col­lec­tive choices for them­selves. The con­straint to indi­vid­uals is intended to leverage the most robust theory in biology: evo­lu­tion through natural selec­tion. If we are to choose a mech­a­nism which is most reliable as an explana­tory factor in biology, that is the one. Thus, for optimum mech­a­nistic robust­ness, Volker and Railsback suggest we focus at the level upon which natural selec­tion acts. Notwithstanding the pos­si­bility that the indi­vidual is not that level (Dawkins 1989), the indi­vidual is cer­tainly the level at which we can most tractably study selec­tion, and thus this is the foun­da­tion they rec­om­mend for devel­oping a new domain science.

Remote Sensing and Ecology

Data chal­lenges in ecology.

Peters is not being alarmist when he asserts that ecology has problems. Most sciences have as their domain iso­lat­able subjects. Physics studies forces and types of matter which retain their prop­er­ties even when in iso­la­tion from each other. Although chem­istry is more fun­da­men­tally about inter­ac­tion of com­po­nents, those inter­ac­tions can be real­is­ti­cally captured in the confined and con­trol­lable space of a test tube. Other branches of biology may be in part con­cerned with the inter­ac­tions of their subjects with the outside world or each other, but gen­er­ally not as the primary focus of research. Ecology on the other hand is specif­i­cally about what happens in the world, and the world is a highly variable and inter­ac­tive place. A phe­nom­enon of interest may be driven by biotic, abiotic, or a com­bi­na­tion of forces. Potential abiotic drivers, such as tem­per­a­ture, humidity, chemical com­po­si­tion, shade or wind, may vary at the scale of meters or less, and some or all may be inter­ac­tively incident on a subject. Biotic drivers– other organisms–are even more abundant and are also poten­tially variable and inter­ac­tive. Establishing any entity as causal with any given effect can be a chal­lenge. There is always the lurking pos­si­bility that a third, unmea­sured factor indi­rectly cor­re­lates what you thought were a causally linked pair. Laboratory studies can limit variance, but at the cost of explana­tory power: just because flour beetles behave a certain way in the lab, doesn’t mean other beetles nec­es­sarily do the same in the crazy real world, even if it is sug­ges­tive of it. For any per­tur­ba­tion or obser­va­tion exper­i­ment, gath­ering enough data to convince yourself that all drivers not under study have been averaged out can be a sig­nif­i­cant chal­lenge. There is often a serious tension between the need for high sample sizes in the face of sub­stan­tial envi­ron­mental variance and the amount of time, resources, hip waders and under­grad­u­ates on hand to gather the data. Time is another sig­nif­i­cant factor. Often eco­log­ical systems cycle over the course of seasons or even decades. Studying a classic predator-​​prey rela­tion­ship such as wolves and moose for instance, may yield appar­ently random results at the scale of years or even gen­er­a­tions, when real cycles do exist and can be observed if data is recorded for many decades. Such long-​​term research programs exist, but are rare due to the obvious resource and per­sonnel chal­lenges they face.

The poten­tial for remote sensing.

Given these fun­da­mental chal­lenges, remote sensing offers some hope as a new, dif­ferent, and poten­tially resource-​​effective source of data. Remote sensing (RS) is very broadly the use of sensors to record char­ac­ter­is­tics of subjects which are not phys­i­cally con­tiguous with the observer. In practice that usually means satel­lite and some­times plane-​​based imaging of land­scapes, although other means exist. In contrast, tra­di­tional “ground-​​sampling” typ­i­cally consists of recording obser­va­tions at a series of precise points over an area of interest. Those obser­va­tions may them­selves be as precise as time and method­ology allows: tree rings may be cored, streams may be chem­i­cally assayed, florets may be counted. But there is almost always more space in between the mea­sure­ments than the mea­sure­ments them­selves rep­re­sent. Quadrats are usually spread out. Remote sensing by contrast measures ubiq­ui­tously but fuzzily. In a remotely sensed image, every place on the ground has a cor­re­sponding pixel. It may however share that pixel with a sub­stan­tial parcel of the sur­rounding ground, depending on the spatial res­o­lu­tion of the sensor. Depending on the sensor there may also be serious lim­i­ta­tions as to what infor­ma­tion has been recorded about that patch of ground. ‘Panchromatic’ sensors can only record levels of visible spectral reflectance of a canopy, not the number of rings in the trees below. ‘Multi-​​spectral’ sensors can record the presence of a stream, but probably not much about its chemical composition.

Even given these data-​​quality lim­i­ta­tions, remote sensing should at the very least rep­re­sent a new type of data, and con­se­quently rep­re­sent a new channel of eco­log­ical explo­ration. Further, many of the dimen­sions of data lim­i­ta­tion are being pushed back, in part by tech­no­log­ical improve­ment in sensors, and largely by improve­ments in data inter­pre­ta­tion. Careful analysis of spectral reflec­tion at the level of the leaf or needle (Stimson at al 2005) may scale to mea­sure­ment of subtle phys­i­o­log­ical status at the level of tree, canopy or land­scape (Panek and Ustin 2004), and inves­ti­ga­tion of the inter­ac­tion of light and fluid may open up streams and seabeds to aerial inspec­tion (Goodman and Ustin2003). Experience with the inte­gra­tion of imagery from multiple sensors to produce common data sets, and the analysis of spatial in com­bi­na­tion with spectral pattern, gen­er­ates syn­thetic res­o­lu­tions pre­vi­ously unre­solv­able (Greenberg et al 2006). Consequently, the trade-​​off of ubiquity of data versus pre­ci­sion of data seems to increas­ingly weigh in favour of remote sensing as an eco­log­ical data-​​gathering tool.

Temporally, remote sensing can cover days, seasons, years, and even in some cases decades. Satellites are typ­i­cally the platform with the best coverage of an area in time, The best may revisit a study site daily, the worst monthly, or only when requested. The longest-​​term eco­log­ical mon­i­toring project that can be con­ducted cur­rently stands at about 40 years, begin­ning in 1963 when the American CORONA spy satel­lites began recording black and white images of the world and dropping film can­is­ters out of low orbit. While this hard limit may at first seem like a mark against RS, it can compare favor­ably with ground-​​sampling, which is limited to the present and future and which becomes expen­sive if even monthly revisits to a study site are required.

Despite these ecological-​​research-​​friendly prop­er­ties, actual use of remote sensing data in eco­log­ical research seems to be limited. Although widely used for mapping and mon­i­toring of con­ser­va­tion concerns (for example, Greenberg et al 2005, Stimson et al 2005b), RS data is much less likely to be used to test eco­log­ical hypotheses. In instances when RS data does tran­scend mere mapping and mon­i­toring, it is typ­i­cally to be incor­po­rated into dynamic modeling, but only to provide a “before” map for the exe­cu­tion of estab­lished spa­tially pre­dic­tive process models. It is less commonly used for the testing of mech­a­nisms that may ulti­mately be built into new pre­dic­tive process models. This may be due to the relative novelty of the tool, or the tra­di­tional pre­dom­i­nance of engi­neers in its use, or the cross-​​disciplinary demands of inte­grating tech­nical knowl­edge with eco­log­ical knowl­edge. As remote sensing data pro­lif­er­ates and improves, and the tech­nical know-​​how to inter­pret it grad­u­ally pro­lif­er­ates and improves, the poten­tial for its use as a driver of eco­log­ical theory seems apparent.

Remote Sensing and Individual-​​Based Ecology

Spatially-​​explicit models are a common feature of IBM, and as with remotely sensed data, that spa­tiality is typ­i­cally of 2 dimen­sions. Although individually-​​based dynamics may alter­na­tively be rep­re­sented as, for instance, spa­tially implicit topolo­gies, a survey of the sample IBMs reported in Grimm and Railsback (2005 ch. 6.2−6.5) yields 10 out of 16 models which rep­re­sent space as a 2 dimen­sional plane with non-​​varying scale. There is an obvious poten­tial for the inter­ac­tion of remotely sensed imagery and the output from such spatially-​​explicit IBMs (or for the use of RS imagery as an input of IBMs). There are extant computer science chal­lenges for the actual inte­gra­tion of RS data into IB models (Brown et al 2005b) but these are in the process of being iden­ti­fied and addressed and are cer­tainly surmountable.

A key issue in the inte­gra­tion of remote sensing and specif­i­cally indi­vidual–based ecology is the capacity to resolve indi­vid­uals. Less chal­lenging but also nec­es­sary is the ability to encom­pass the system which their aggre­gate behavior is expected to char­ac­terize. Application of sophis­ti­cated remote sensing tech­niques may well be nec­es­sary to process and produce spatial scenes for incor­po­ra­tion or com­par­ison with spatial IBMs.

More gen­er­ally, the impor­tance of spatial pattern in ecology is well estab­lished, and spatial ecology is a growing field. If IBMs can be made more spa­tially robust, either through the tem­pering effect of empir­ical data testing or other methods, they have the poten­tial to shape and be shaped by an exciting and dynamic sub­dis­ci­pline of ecology (Grimm et al 2005, With and King 2004, Camarero et al 2005, Brown et al 2006, Watt 1947, Levin 1992)

Related to the ‘natural history’ v. ‘hard science’ divide in ecology is the divide between field ecol­o­gists and the­o­rists. It may not be far from factual to stereo­type field ecol­o­gists as dis­missing the­o­rists as only vaguely relevant and vice a versa. This divide may in part be related to the dif­fi­culty of col­lecting sta­tis­ti­cally sig­nif­i­cant quan­ti­ties of data. Grand theory tends to remain grand and somewhat remote in the absence of ready testa­bility. A platonic air of spec­u­la­tion founded only on better-​​established spec­u­la­tion seems to pervade the­o­ret­ical biology. As it is so hard to actually check if a grand-​​scope eco­log­ical theory is true, math­e­mat­ical elegance eclipses worldly verisimil­i­tude as the highest prac­tical attain­ment. Individual-​​based ecology should not be allowed to drift in this direc­tion. While the spir­i­tual and aes­thetic plea­sures of math­e­mat­ical theory may be real and positive, the demands of con­ser­va­tion planning are too pressing to allow abstract elegance as a priority for the dis­ci­pline. The avail­ability of real data in a form neatly applic­able to the models would help to keep the focus on mech­a­nistic rep­re­sen­ta­tion and accurate prediction.

To this end, there is a clear path for inte­gra­tion of remotely-​​sensed data into IBE. Although non­lin­ear­i­ties may render it unlikely for a model to repro­duce the exact patterns of a real land­scape, they may well be able to repro­duce the more general patterns. Grimm and Railsback devote a chapter (2005 ch. 3) to the methods and goals of pattern-​​oriented modeling. Although they don’t dwell sub­stan­tially on methods for gath­ering data for pattern-​​based com­par­ison to model pre­dic­tions (hint: remote sensing) they do point out that by refining IBMs based on com­par­isons of their output to observed spatial data, individual-​​based modeling can become a com­po­nent of the cycle of observation-​​hypothesis-​​testing-​​observation gen­er­ally asso­ci­ated with the induc­tive sci­en­tific method.

Sensors, Systems, and Scales

Choosing a sensor and a system.

Commonly, remote-​​sensing research programs are guided by the research lineage of the indi­vidual doing the research. Choice of sensor and/​or system is typ­i­cally the con­se­quence of the sensors and systems the researcher’s lab­o­ra­tory col­leagues have pursued pre­vi­ously. This kind of path-​​dependence in research program is not nec­es­sarily a bad thing: the deeper and more subtle a researcher’s under­standing of the usage tricks of a given imagery type, and the deeper their under­standing of the bio­log­ical real­i­ties they are attempting to sense, the more likely their efforts will bear real fruits. In any case, often the rela­tion­ships estab­lished between data providers and data users make it such that a given type of imagery is most easily attain­able for a given user.
Given the laundry-​​list of hurdles to the remote sensor trying to resolve indi­vid­uals however (see below), such a researcher may have to be willing to cast a wider net. Matching a sensor and a system will be a victory in itself, and it may be nec­es­sary to submit oneself to serendipity of scale overlap, regard­less of pref­er­ence or expe­ri­ence. Consequently, a rigorous cat­a­loguing of the relevant scales of sensors and systems may rep­re­sent a useful or even nec­es­sary pre­cursor to choosing a specific research topic (Stimson 2007).

Challenges: spatial resolution.

Probably the single greatest chal­lenge facing the individual-​​based remote sensor is the ability to resolve the spatial dynamics of indi­vid­uals. Given that the maximum current res­o­lu­tion of satel­lites is around the 1/​2m scale (and that is only avail­able for the simplest, panchro­matic imagery), and to really resolve a subject the subject must be larger than the minimum pixel size, the vast majority of species on earth are nec­es­sarily pre­cluded from con­ven­tional remote sensing. There are at least two responses to this problem: one is to accept that satel­lites can only be used for medium and large scale veg­e­ta­tion (e.g. trees and bushes—groundcover is also resolv­able at this scale but not at the truly indi­vidual level). Another is to use plane-​​based sensors, in par­tic­ular aerial pho­tog­raphy, which may have res­o­lu­tions as low as cen­time­ters, but which suffer from high cost and low repeata­bility. These would make smaller scale veg­e­ta­tion and sessile organ­isms a possible subject. A third is to con­struct purpose-​​built instru­ments, which have a high front-​​end cost but poten­tially lower incre­mental costs, at least compared to plane-​​based instru­ments, and which are less easily repeat­able than satel­lites but more so than plane-​​based sensors. Unlike satel­lites (and to an extent plane-​​based sensors) they cannot collect data in the past. Like plane-​​based sensors, purpose-​​built sensors provide the pos­si­bility of smaller-​​scale veg­e­ta­tion and sessile organ­isms (e.g. Guichard et al 2000 and 2003, Guichard 2005).

Given that spatial res­o­lu­tion seems to make medium-​​and-​​large scale veg­e­ta­tion the most likely subject for individual-​​based remote sensing (‘IB-​​RS’), it is encour­aging to note that effort and skill is being deployed in using hybrid image pro­cessing tech­niques to resolve satel­lite imagery to the level of indi­vidual tree crowns (e.g. Greenberg et al 2006). This rep­re­sents a poten­tial ‘prime target’ for IB-​​RS. Consider, for example, the data needs of Solé and Manrubia 1995. Another con­sid­er­a­tion is a focus on ‘high-​​contrast’ indi­vid­uals, those which are more easily resolv­able due to a high dif­fer­en­tial of their reflec­tive char­ac­ter­is­tics from the back­ground matrix they exist in. These are commonly organ­isms in biot­i­cally stressed envi­ron­ments, such as krummholtz trees at the tree line (Alftine and Malanson 2004) or woody shrubs and trees in semi-​​arid ecosys­tems (Stimson et al 2005). Such systems tend to have biotically-​​driven dynamics which are not nec­es­sarily gen­er­al­iz­able to more common abiotically-​​constrained systems. One of my personal research interest is in the relative preva­lence of positive feedback in biot­i­cally con­strained systems and negative feedback in abi­ot­i­cally con­strained systems, and as such these high-​​contrast systems are of interest to me.

Challenges: temporal resolution.

Temporal res­o­lu­tion inter­acts with spatial res­o­lu­tion in pro­viding some of the hardest con­straints on individual-​​based sensing. No satel­lite and, to my knowl­edge, no other existing sensor observes any given area con­tin­u­ously. Without con­tin­uous obser­va­tion, and with the excep­tion of very large indi­vid­uals indeed, it would be impos­sible to identify indi­vid­uals from time sample to time sample except with regards to their position. So any inves­ti­ga­tion must be of dynamics which occur on a temporal scale equal or greater than the revis­iting time of the sensor. This would seem to preclude any non-​​sessile animals (except perhaps sloths) and place the focus purely on veg­e­ta­tion. The only excep­tions might be for purpose-​​built, per­son­ally deployed sensors. Another possible avenue of progress (not much con­sid­ered here due to my lack of famil­iarity with it) is the rapid progress being made in widely deployed ‘spa­tially embedded sensor networks’ of the ‘smart dust’ variety (for these and other exciting fron­tiers, see Green et al 2005). Perhaps some of those sensors could be image-​​based, poten­tially pro­viding coverage of a medium or small-​​scale land­scape nearly ubiq­ui­tously in time and at fine spatial resolution.

Challenges: temporal scope.

One of the won­derful things about satel­lites is that they’ve already col­lected the data. Although it is possible to acquire the freshest data, it is equally easy to use data from any past season. Many (though not all) satel­lites record con­stantly and globally and archive com­pre­hen­sively. Consequently, if a match can be found between sensor and system, then the data is a done deal. Field work will inevitably be required, but research is con­tin­gent more on the known and doc­u­mented past sensing pro­gramme than its future funding and main­te­nance. The chal­lenge is not so much to collect the data as to make it relevant.
This is not so true of non-​​satellite RS data. In some cases, the com­mis­sioners of expen­sive plane-​​collected imagery will be reluc­tantly willing to release it for further use, but the chances of the loca­tions and times of previous research being relevant to new research are reduced compared to the more com­pre­hen­sive back-​​catalogs of satel­lite data. Individually built and deployed sensors, to the extent that they exist (e.g. Guichard et al 2000), offer a slightly dif­ferent set of chal­lenges again. In these cases, there will be a very limited set of back-​​cataloged data avail­able. Individual will­ing­ness to release that data will likely depend on indi­vidual idio­syn­crasy. Perhaps more impor­tantly, the chance that a prospec­tive researcher will know that someone else’s individually-​​collected data exists is even less than for plane– or satellite-​​based col­lec­tion. The campaign his­to­ries of research planes are often doc­u­mented in some public way, and satel­lite data typ­i­cally has elab­o­rate data access tools on their respec­tive websites. Individually-​​collected imagery is pri­marily dis­cov­ered through indi­vidual papers in the lit­er­a­ture or through personal acquaintance.

Challenges: resolv­able system characteristics.

Many IBMs focus sub­stan­tially on presence/​absence, or state vari­ables deriv­able from repeated presence/​absence mea­sure­ment, such as age (Greenberg et al 2005) . For these char­ac­ters, the limiting sensor char­ac­ter­is­tics are likely spatial and temporal res­o­lu­tion. However, there is sub­stan­tial pos­si­bility of gath­ering more subtle, even phys­i­o­log­ical mea­sure­ment using sensed land­scape imagery (Panek and Ustin 2004), and these char­ac­ter­is­tics might be used as inputs for or tests of anal­o­gous state vari­ables in IBMs. In such cases, the imaging char­ac­ter­is­tics of sensors, such as multi-​​spectral, hyper-​​spectral, LIDAR, and radar, become sub­stan­tial factors. Although these sensor-​​characteristic concerns rep­re­sent the bulk of what worries remote sensors, they would be highly study-​​dependant, and for IB-​​RS would likely become a limiting factor only once less subtle con­straints of scale have been resolved.

Challenges: cost and access.

The finan­cial and pro­ce­dural burden of obtaining remotely-​​sensed imagery varies wildly depending on the public or private nature of the sensing platform, and the personal con­nec­tions of the remote sensor. Although dif­fi­cult to char­ac­terize a priori these logis­tical con­sid­er­a­tions are likely to frus­trate any non-​​pecuniary assess­ment of the ideal sensor-​​system match.

Challenges: spatial scope.

Most sensors rep­re­sent a likely match between spatial res­o­lu­tion and spatial scope. The spatial scope of purpose-​​built and airborne sensors tends to be from 100s of meters to kilo­me­ters. Those scales seem likely to encom­pass the dynamics of most systems which I can imagine being resolved at airborne spatial res­o­lu­tions, which is typ­i­cally meters to dozens of meters. Lower res­o­lu­tion, satellite-​​based sensors typ­i­cally encom­pass scenes from many to 100s or even 1000s of kilo­me­ters across. For systems of greater spatial scope, mosaicing of con­tiguous images rep­re­sents greater access costs and poten­tial image-​​processing road-​​blocks, but is often still a viable option. Spatial scope is thus a mer­ci­fully minor chal­lenge for IB-​​RS.

Case Studies

Case study one: treeline development.

Alftine and Malanson (2004) con­ducted an inves­ti­ga­tion of the patterns of tree estab­lish­ment at the alpine-​​subalpine tran­si­tion zone. They observed that at the treeline edge, the pattern of tree growth was patchy and irreg­ular. Noting that “large-​​scale prop­er­ties of veg­e­ta­tion are inex­tri­cably linked with the fine-​​scale mech­a­nisms by which indi­vidual plants can affect and respond to their local envi­ron­ment”, they devel­oped a model of tree estab­lish­ment built on a cellular automata frame­work, in which per-​​grid-​​cell tree estab­lish­ment was affected by wind exposure. According to this model, the like­li­hood of a tree growing in a location was increased if other trees were already in place around that location to block the pre­vailing wind, and enhance the soil quality. Field sampling was con­ducted to measure 2-​​dimensional tree presence and absence over plots mir­roring their model grids in scale. The model, which rep­re­sented inter­ac­tions at the indi­vidual level, was tested against a null model of non-​​interactive tree estab­lish­ment for its ability to match the observed data on a few simple pattern metrics. The IB version of the model was sig­nif­i­cantly more real­istic in the spatial patterns of tree estab­lish­ment produced.

This rep­re­sents an example of an IBM study which suc­cess­fully incor­po­rated empir­ical data to test and refine its assump­tions. It also rep­re­sents an example that could have more easily con­ducted with remotely sensed, or greatly expanded. Although the cost of tree-​​resolution data can be pro­hib­i­tive, the lack of neces­sity for a time series of data means that such a study could easily ‘piggy-​​back’ on existing data from any existing study. Many high-​​resolution scenes have been recorded at the alpine-​​subalpine tran­si­tion and could be borrowed or begged from fellow researchers (see 4.f). Of the vari­ables field-​​sampled for the study, only two (tree position and ele­va­tion) were col­lected across the entire sample area. Both vari­ables could have been readily gen­er­ated using high-​​resolution imagery. All other state vari­ables in the model (wind direc­tion, soil quality, initial site quality) were devel­oped by point sampling and inter­po­la­tion based on ele­va­tion. Point sampling in a similar number of loca­tions could be part of a RS-​​based version of this study, and would rep­re­sent an excuse for field work.

Case study two: inter­tidal dynamics.

Guichard and col­leagues have con­ducted a series of studies around the dynamics of growth and survival of mussels and other inter-​​tidal organ­isms. The research has focused on the inter­ac­tion of biotic with abiotic forces (such as tidal flow and bed topog­raphy) and has included con­sid­er­able study of individual-​​level (Guichard 2003) or other complex inter­ac­tions (Guichard et al 2005). To build on the empir­ical nature of this work, a remote sensing platform was con­structed using a 6m hydrogen blimp and sus­pended consumer camera , (Guichard et al 2000). By taking stereo images multiple times over a season, a high res­o­lu­tion 3D map of veg­e­ta­tion presence and devel­op­ment was con­structed. The initial study focused on rela­tion­ships between topo­graphic het­ero­geneity and organism density, and explic­itly compared model pre­dic­tions with observed patterns. With the remote sensing platform and asso­ci­ated protocol estab­lished, the authors are well-​​positioned to develop their models of the rela­tion­ships of scale and inter­ac­tion with more robust empir­ical data.

This rep­re­sents an example of the ideal inte­gra­tion of remote sensing with theory: a well estab­lished system is being studied using a sensor capable of cap­turing the relevant scales of both indi­vid­uals and the encom­passing system. Of par­tic­ular interest is that fact this study includes mussels as study organ­isms: by focusing on a sessile species, this rep­re­sents remote sensing of non-​​vegetation, probably the only example I am aware of (or can imagine). Although limited by the labor nec­es­sary to deploy the platform in the field for each desired time-​​sequence, the fit of sensor res­o­lu­tion to system scale yields con­sid­er­able latitude for eco­log­ical exploration.

References.

Alftine KJ, Malanson GP. 2004. Directional positive feedback and pattern at an alpine tree line Journal of Vegetation Science. 15.
Batty M, Torrens PM. 2001.Modeling com­plexity: the limits to pre­dic­tion. Cybergeo. 201.
Brown DB, Page S, Riolo R, Zellner M, Rand W. 2005. Path depen­dence and the val­i­da­tion of agent-​​based spatial models of land use. International Journal of Geographical Information Science. 19(2).
Brown DG, Riolo R, Robinson DT, North M, Rand W. 2005. Spatial process and data models: Toward inte­gra­tion of agent-​​based models and GIS.Journal Geographical Systems..7:1(23).
Brown DG, Aspinall R, Bennett DA. 2006. Landscape models and expla­na­tion in land­scape ecology—a space for gen­er­a­tive land­scape science? The ProfessionalGeographer. 58(4).
Camarero JJ, Gutiérrez E, Fortin MJ, Ribbens E. 2005. Spatial patterns of tree recruit­ment in a relict pop­u­la­tion of Pinus uncinata: forest expan­sion through strat­i­fied dif­fu­sion. Journal of Biogeography. 32.
Clarke KC, Gaydos J, Hoppen S. 1997. A self-​​modifying cellular automaton model of his­tor­ical urban­iza­tion in the San Francisco Bay area. Environment and Planning B: Planning and Design. 24.
Dawkins R. 1989. The Selfish Gene, Oxford University Press.
DeAngelis DL, Gross LJ. 1992. Individual-​​based models and approaches in ecology: pop­u­la­tions, com­mu­ni­ties and ecosys­tems. Chapman & Hall.
Goodman JA, Ustin SL. 2003. Airborne hyper­spec­tral analysis of coral ecosys­tems in the Hawaiian Islands. 30^th International Symposium on Remote Sensing of Enivronment. Florida,USA.
Green JL, Hastings A, Arzberger P, Ayala FJ, Cottingham KL, Cuddington K, Davis F, Dunne JA, Fortin MJ, Gerber L, Neubert M. 2005.
Complexity in ecology and con­ser­va­tion: Mathematical,
sta­tis­tical, and com­pu­ta­tional chal­lenges. Bioscience. 55(6).
Greenberg JA, SC Kefauver SC, Stimson HC, Yeaton CJ, Ustin SL. 2005. Survival analysis of a neotrop­ical rain­forest using mul­ti­tem­poral satel­lite imagery. Remote Sensing of Environment. 96 (2).
Greenberg JA, Dobrowksi SZ, Ramirez CM, Tuil JL, Ustin SL. 2006. A bottom-​​up approach to veg­e­ta­tion mapping of the Lake Tahoe basin using hyper­spa­tial image analysis. Photogrammetric Engineering & Remote Sensing. 72 (5).

Grimm, .1999. Ten years of individual-​​based mod­el­ling in ecology: what have we learned and what could we learn in the future? Ecological Modelling 115
Grimm V, Railsback SF. 2005. Individual-​​based modeling and ecology. Princeton University Press.
Guichard F, Bourget E, Agnard JP. 2000. High-​​resolution remote sensing of inter­tidal ecosys­tems: A low-​​cost tech­nique to link scale-​​dependent patterns and processes. Limnology and Oceanography., 45(2).
Guichard F, Halpin PM, Allison GW, Lubchenco J, Menge BA. 2003. Mussel dis­tur­bance dynamics: Signatures of oceano­graphic forcing from local inter­ac­tions. The American Naturalist. 161(6).
Guichard F. 2005. Interaction strength and extinc­tion risk in a meta­com­mu­nity. Proceedings of the Royal Society B. 272(1572).
Hirsch AE, Gordon DM. 2001. Distributed problem solving in social insects. Annals of Mathematics and Artificial Intelligence. 2001.
Hartvigsen G, Kinzig A, Peterson G. 1998. Use and Analysis of Complex Adaptive Systems in Ecosystem Science: Overview of Special Section. Ecosystems. 1.
Levin SA. 1992. The Problem of Pattern and Scale in Ecology. Ecology. 73:6.
Levin, S. 1999. Fragile Dominion: Complexity and the Commons. Perseus Books.
Levin SA. 1999. Complex adaptive systems: Exploring the known, the unkown and the unknow­able. Bulletin (New Series) of the American Mathematical Society. 40(1).
Lotka AJ. 1932. The growth of mixed pop­u­la­tions: Two species com­peting for a common food supply. Trans. Scudo FM, Ziegler JR. 1978. The golden age of the­o­ret­ical ecology. Springer-​​Verlag.

Maytin CE, Acevedo ME, Jaimez R, Andressen R, Harwell M, Robock A, Azocar 1995. A. Potential Effects of Global Climactic Change on the Phenology and Yield of Maize in Venezuela. Climatic Change 29.
Moloney KA, Levin SA. 1996. The effects of dis­tur­bance archi­tec­ture on landscape-​​level pop­u­la­tions dynamics. Ecology 77(2).
Nuetel AM, Heesterbeek JAP, Ruiter PC. 2002. Stability in real food webs: Weak links in long loops. Science. 296.
Olaf KN, Visintini F. 2005, Her Majesty the Queen in Right of Canada. 2005. Assessment of poten­tial for remote sensing detec­tion of bark beetle-​​infested areas during green attack: a lit­er­a­ture review. Mountain Pine Beetle Initiative Working Paper 2005–2. Natural Resources Canada.
Panek J, Ustin SL. 2004. Ozone uptake in relation to water avail­ability in pon­derosa pine forest: mea­sure­ments, modeling and remote-​​sensing. National Parks Service Final Report PMIS #76735. Accessed Mar 2007: http://​www2​.nature​.nps​.gov/​a​i​r​/​P​u​b​s​/​p​d​f​/​t​o​x​i​c​s​/​P​a​n​e​k​R​e​p​o​r​t​2​0​0​4​.​pdf
Peter, RH. 1991. A critique for ecology. Cambridge University Press.
Solé R, Manrubia SC. 1995. Are rain­forests self-​​organized in a critical state? Journal of Theoretical Biology. 173.
Solé RV, Gamarra JGP, Ginovart M, López D. 1999. Controlling chaos in ecology: from deter­min­istic to individual-​​based models. Bulletin of Mathematical Biology. 61.
Stimson HC, Breshears DD, Kefauver SC, Ustin SL. 2005. Spectral sensing of foliar water con­di­tions in two co-​​occurring conifer species: Pinus edulis and Juniperus monosperma. Remote Sensing of Environment. 96(1).
Stimson H, Songer M, and Leimgruber P. 2005b. Mapping South Asian dry dipte­ro­carp forest using dual season imagery. 2^nd Smithsonian Geographic Information Systems (GIS) Conference. Febuary 11. Washington, DC, USA.
Stimson, HC. 2007. Proposal: A remote-​​sensor/​ecological-​​system auto-​​matcher. http://​hugh​stimson​.org/​b​l​o​g​/​2​0​0​7​/​0​3​/​1​6​/​p​r​o​p​o​s​a​l​-​a​-​r​e​m​o​t​e​-​s​e​n​s​o​r​-​e​c​o​l​o​g​i​c​a​l​-​s​y​s​t​e​m​-​a​u​t​o​-​m​a​t​c​h​er/

Van der Maarel E. 1996. Pattern and process in the plant com­mu­nity: fifty years after A.S. Watt. Journal of Vegetation Science. 17(1).
Volker G, Revilla E, Berger U, Jeltsch F, Mooj WM, Railback SF, Thulke HH, Weiner J, Wiengand T, DeAngelis D. 2005. Pattern-​​Oriented Modeling of Agent-​​Based Complex Systems: Lessons from Ecology Science. 310.
Watt AS. 1947. Pattern and Process in the Plant Community. The Journal of Ecology. 35.
With KA, King AW. 2004. The effect of land­scape struc­ture on com­mu­nity self-​​organization and critical bio­di­ver­sity Ecological Modelling. 179.
Volterra, V. 1927. Variations and fluc­tu­a­tions in the numbers of coex­isting animal species. Trans. Scudo FM, Ziegler JR. 1978. The golden age of the­o­ret­ical ecology. Springer-​​Verlag.