دانلود کتاب راهنمای هوش مصنوعی زمین فضایی

It’s a great honor for me to be asked to contribute a foreword to this Handbook of Geospatial Artificial Intelligence. GeoAI has very quickly become a major topic of new research and development in geography, geographic information science (GI- Science), and in many of the disciplines that concern themselves with the complex patterns and processes that can be found in the geographic domain (that is, the sur- face and near-surface of the Earth). To this old dog, it represents not only an exciting set of new tricks, but a reinvigoration of the old field of geographic science in di- rections that are fundamentally different from what came before. It benefits from a perfect storm of trends: the availability of many new sources of data, from remote sensing, social media, and sensor networks; access to almost unlimited resources of computational power; and the emergence of powerful new methods of data analysis and machine learning. More fundamentally, GeoAI reflects a radical shift in our approach to under- standing the geographic domain. Half a century ago the geographic sciences modeled themselves on the senior sciences of physics and chemistry in searching for universal principles (Bunge, 1966). These principles should apply everywhere and at all times, just like Mendeleev’s periodic table of the elements. They should also be simple, in accordance with the principle known as Occam’s Razor: when two hypotheses might explain the same phenomenon, one should adopt the simpler. The use of Newton’s law of gravitation to explain how human communications tend to diminish with in- creasing distance provides an excellent example, and led to extensive research on the modeling of social interaction. The fact that such models will never provide perfect fits to actual geographic data was inconvenient, but their estimates were nevertheless fit for use in a host of applications. Moreover, the models provided a norm or stan- dard that would be helpful in identifying exceptions, anomalies, and special cases. By the mid-1980s, however, this attempt to pursue geographic science by emu- lating physics had run its course. The geographic world was clearly too complex for a set of mechanistic explanations, and more powerful techniques would be needed if we were to discover patterns and identify processes. Some pursued techniques that abandoned universality by adopting what we have come to call place-based meth- ods, which allow explanations to vary across space or time, or both (e.g., Geograph- ically Weighted Regression; Fotheringham, Brunsdon, and Charlton, 2002). Others began to build analysis engines that would explore entire sets of models rather than a few narrowly defined hypotheses. Openshaw, for example, built a series of what he termed geographical analysis machines that would give the data an increasing role in driving the model selection process; much later this approach became enshrined in the principle of the Fourth Paradigm, “Let the data speak for themselves” (Hey, Tans- ley, and Tolle, 2009). Openshaw was an early user of the term artificial intelligence, and his ideas were collected in an appropriately titled but prescient book that was published shortly before the turn of the century (Openshaw and Openshaw, 1997). In a similar vein, Dobson began writing about what he called automated geography (Dobson, 1983); both he and Openshaw argued that the computer should become increasingly engaged in the research process. Several decades later these ideas are entering the mainstream of the geographic sciences, but with an important difference. While both Dobson and Openshaw were trained in the domain-specific techniques of geographic analysis, today’s methods of machine learning draw from no particular domain science, but instead apply basic approaches that are essentially the same whatever the subject matter. Thus one of the most urgent needs in GeoAI is for techniques that incorporate the general principles that we know to be true of the geographic domain: spatial dependence, spatial het- erogeneity, scaling, etc. But while they may appear to be neutral, applying equally well to any domain, techniques such as neural networks and their more recent devel- opments such as DCNN (deep convolutional neural networks) may to some extent emulate rudimentary ideas about the workings of the human brain and its instinctive search for patterns. It seems somewhat ironic that in rejecting the simple mechanistic models of classical physics, we have gravitated to the vastly more complex but sim- ilarly mechanistic world of neural networks. Moreover, as some of the chapters of this collection show, the use of DCNN to analyze imagery explicitly invokes the geo- graphic concept of spatial dependence, or what we know familiarly as Tobler’s First Law of Geography.