AI and The Environment
AI Blueprints for 16 Environmental Projects Pioneering Sustainability
By Cindy Mason

ai and environment by cindy
        mason cover
This Book in A Nutshell

If You Don’t Have Time to Read The Whole Book

Chapter 1  Fire Fighting
Combining Human Assessment and Reasoning Aids for Decision-making in Planning Forest Fire Fighting
This chapter addresses fire fighting resource planning with a hybrid AI system for automatically planning first attacks to a forest fire based on work organization in an Italian Provincial center. The complexity of fire fighting is typical of environmental problems.  Fire is a dynamic phenomenon that changes and whose evolution is determined by weather conditions like wind direction and intensity, by humidity, and by fuel type, which changes rapidly in sometimes unpredictable ways, requiring fast decisions.  Data about these operating conditions are often uncertain, incomplete, and in some cases totally absent.  The decision making automation is complicated with relevant fire events evolving with different time and spatial scales.  The planning problem of a fire emergency, like many environmental emergencies, is complicated  because multiple organizations are involved with decision making over the fire territory and cooperation is needed for good decision making on the strategies to fight the fire: which  fire front, where to locate resources, what needs most attention (e.g. railway, houses, etc.). There are also decisions about which order to do things.  In this way past knowledge is very important.  Their approach uses multiple AI techniques: lazy learning, case-based reasoning and constraint reasoning. The system is aimed at supporting the user in the whole process of forest fires management and the user always remains active in the ultimate decisions.

This work conducted at the IRST in Italy by Paolo Avesani, Francesco Ricci, and Anna Perini.

Chapter 2 Flood Prediction
Introducing Boundary Conditions in Semi-quantitative Simulation
The chapter addresses flood prediction and water supply control and describes a hybrid AI  method that addresses the problem of incomplete and  imprecise information that generally plagues many environmental simulation systems. Predictions with standard numerical simulations for boundary value problems* can be error prone because they require precision but in reality the information is imprecise and incomplete. For example, when flood conditions approach, empirical data on the level/flow-rate curve for rivers becomes less and less accurate.  In general, the precise shape and size of a body of water is rarely known. The task of flood control and water supply prediction is both difficult and vitally important.  For example, a lake has a dam with floodgates that can be opened or closed to regulate the water flow through power generating turbines, the water level (stage) of the lake, and the downstream flow. The goal of a controller is to provide adequate reservoir capacity for power generation, consumption, industrial use, and recreation, as well as downstream flow. In exceptional circumstances, the controller must also work to minimize or avoid flooding both above and below the dam. The conceptual and practical aspects addressed by the AI system include the ontology (actions vs. measurements), the temporal scale (instantaneous vs. extended changes), the impact of discontinuity on model structure and the consequences of incompleteness in predictions.  Environmental simulation tools are useful to careful evaluate the effect of actions in critical and dynamically changing situations.  They evaluate empirically derived models and parameters,  and help to forewarn of undesired possible future situations. The hybrid AI simulation method extends the application of qualitative AI modelling methods to include simulating dynamic systems and problem of handling boundary conditions*.  

*in a simulation system, boundary value problems specifying how external influences on dynamic systems vary over time

This work conducted at the Università de Udine in Italy by Georgio Brajnik.

Chapter 3   Sewage and Pollution
Integrating General Expert Knowledge and Specific Experimental Knowledge in Waste Water Treatment Plant
The chapter addresses pollution level control for waste water treatment plants using a hybrid AI system that combines both learning from past experience and from domain knowledge for real-time control of a wastewater treatment plant (WWTP).The main goal of a wastewater treatment plant is to reduce the pollution level of the wastewater at the lowest cost, that is, to remove, within die possible measure, strange compounds (pollutants) of the inflow water to the plant prior to discharge to the environment. So, the effluent water has the lower levels of pollutants as possible (in any case, lower than the maximum ones allowed by the law).The plants taken as models -in this study- are based on the main biological technology usually applied: the activated sludge process. The target wastewater plant studied is located in Manresa, near Barcelona (Catalonia). This plant receives about 30000 m3/day inflow from 75000 inhabitants. The automated solution to this real-time control problem is a multi-paradigm reasoning architecture able to input and process the different elements of the knowledge learning process, to learn from past experience (specific experimental knowledge) and to acquire the domain knowledge (general expert knowledge).  These are the key problems in real-time control AI systems design. These problems increase when the process belongs to an ill structured domain and it is composed by several complex operational units . Therefore, an integrated AI methodology which combines both learning from past experience and from domain knowledge is proposed. This multi paradigm reasoning provides the target system, a wastewater treatment plant (WWTP), with some advantages over other approaches applied to real world systems.  Due to the dynamic learning environment, the system is able to adapt itself to different waste water treatment plants, making the system to be exportable to any plant with some minor changes. It is only needed to fill the Case library with an initial set of specific cases (operating situations of the concrete WWTP), which can be obtained semi-automatically from real operational data.  All these facts make it more powerful than other single technologies applied to wastewater treatment plants as knowledge-based approaches, statistical process control techniques, fuzzy controller methods, etc., as well as to other complex ill-structured domains. With this approach, the plant can be controlled in normal situations (mathematical control), in abnormal usual situations (expert control) and in abnormal unusual situations (experimental control).

The chapter is based on work conducted in Spain at the Technical University of Catalonia by Miquel Sànchez, Ulises Cortés, and by Ignasi R. Roda, and Manel Poch also in Spain at Universitat de Girona.

Chapter 4   Sustainable Forests/Timber Harvesting
Planning with Agents in Intelligent Data Management for Forestry
The chapter concerns the sustainability of our forests. The largest vegetation on earth are the forests and we place a great demand on them to provide us with wood products - industrially and as consumers. To make sustainable timber harvesting decisions we need to know the state of growth and health of forests over large areas of the earth is essential. The data comes in at about 1Tb/day, is format and media diverse, exists on multiple computing platforms with different access and use policies and includes topographic, soils, hydrology, geology, remote sensing and forest cover descriptions over large areas of the earth.  The AI system helps manage this hairy data problem.  It is used in support of human decision making process by automatically monitoring data and detecting changes and trends on the state of the biosphere and vegetation. The hybrid-AI system uses software agents that combine both learning from past experience and from knowledge. One of the main problems that the AI system must tackle is the update of forest cover maps stored in digital form in geographical information systems (GIS) by processing remotely sensed imagery in order to detect changes in the state of the forest. The agents learn to do this automatically through the use of a training interface that allows human experts to describe how each task is performed.  It is therefore not required to hand code the agents since they are automatically generated by the training interface.These agents can each give a description of the task they perform. These descriptions are then used by a problem solving system that integrates the use of search based planning, case-based reasoning, derivational analogy and machine learning.  The software agent systems learn by unobtrusively observing the manner in which they are used, adapt to the tasks for which they are used, and learn from the circumstances of their use.

This work conducted in Canada by David G. Goodenough at the Pacific Forestry Center and by Daniel Charlebois and Stan Matwin also in Canada at the University of Ottawa.

Chapter 5  Water Pollution Prediction
Water Pollution Prediction With Evolutionary Neural Trees
This chapter addresses water pollution prediction using an evolutionary neural learning method for time series data.  The task studied here is to predict nitrate levels a week ahead in the watersheds of the Sangamon River in Illinois, USA, from the previous values.  The AI method of evolutionary learning networks is generally used for the modeling and prediction of complex systems. In contrast to conventional neural learning methods, genetic learning makes relatively few assumptions about the models of data.  The method is effective in identifying important structures and variables in systems whose functional structures are unknown or ill-defined.  It uses tree-structured neural networks whose node type, weight, size and topology are dynamically adapted by genetic algorithms. Since the genetic algorithm used for training does not require error derivatives, a wide range of neural models can be identified. The performance results compare favorable to those achieved by well-engineered, conventional system-identification methods.  The original study here also aims at giving some indication of the biochemical and physical relationships among the variables and of the controllability of the system.  Application areas for this approach include but are not limited to prediction, monitoring, and diagnosis of complex systems, such as environmental processes.

The chapter is based on work conducted in Germany at the German National Research Center for Computer Science by pioneers Byoung-Tak Zhang, Peter Ohm, and Heinz Mühlenbein.   

Chapter 6  Toxic Algae Blooms
A Qualitative Modeling Approach to Algal Bloom Prediction
This AI project concerns the problem of toxic algae blooms and is a collaboration between researchers from Brazil, Germany and France.  The approach to AI system development is an intelligent model-based systems to support decision making concerning many environmental factors of algae bloom. It is discussed in the context of an algae bloom in the Rio Guaíba in Southern Brazil.  Knowledge-based systems support analysis and decision making using a representation of our human knowledge about the involved algae bloom processes. Because of the very nature of these algae and bloom processes, our knowledge about them, and the information available, this is a great challenge for standard qualitative modeling. This chapter presents preliminary results of our work including: using AI-modeling for the phenomena of the algal bloom and an intelligent process-oriented description of some of the essential mechanisms contributing to algal bloom. In particular, two problems have to be addressed that are typical for modeling ecological systems. First, the spatial distribution of parameters and processes relevant to algae blooms has to be taken into account which leads us to locate processes in or between water body compartments, the elements of a topological partitioning of the area. Second, the various processes involved in an algae bloom development act with speeds of different orders of magnitude (e.g. chemical reactions vs. changes in fish population), which requires AI techniques of time-scale abstraction.  Our approach to modeling the interactions involved in this bloom phenomenon use a language called QPC.  QPC allows the automatic direct expression and representation of physical models of compartments and their interaction, and the application of time-scale abstraction in the composition of a scenario model.

This work was carried out in Brazil, France and Germany with pioneers Waldir Roque at the Federal University of Rio Grande do Sul, Ulrich Heller and Peter Struss from  Technical University of Munich, and Francois Guerrin at INRA Toulouse.

Chapter 7  Recycling and Resource Use in Product Life Cycle
The Green Browser
The chapter is about revealing product life cycle information from the raw material stage through use and eventual disposal or recycling. Unlike a general purpose browser, the Green Browser uses AI methods to focus selectively on environmental (green) product information extracted from the net. Applying AI methods to general browser technology allows us to quickly reveal products bearing positively on green production and environmental protection.  This creates greater public literacy and market selection through a product’s potential impacts: from resource usage and extraction to disposal and dispersal.  A focused browser finds green product information faster than a normal browser.  Faster, easier access to such information supports informed corporate and public decisions and enables stakeholders (e.g., employees, shareholders, consumers, regulators, NGO. etc.) to have automated focused access to all available environmental information. The AI information representation and design schemes proposed for this purpose are called green life cycle model and green life cycle design.   The chapter proposes a representational scheme called green life cycle model which organizes corporate information for the Green Browser. For the purpose of supporting design for life cycle of green products (green life cycle design), the scheme is built to illustrate a product’s potential impacts from the raw material stage through use and eventual disposal or recycling. Firms are encouraged to process their firm-specific information based on the scheme. Second, the chapter discusses how the Green Browser can support information sharing to enable stakeholders to obtain the detailed picture of products.

This chapter is based on work in Japan, with pioneering scientists from the University of Tokyo: Yasushi Umeda, Tetsuo Tomiyama, Takashi Kiriyama, and Yasunori Baba.

Chapter 8  People arguing and making decisions
Support for Argumentation in Natural Resource Management
In this chapter AI helps resolve arguments about natural resources among differently interested parties.  When decisions to be made involve changes to natural resources such as oceans, forests or the atmosphere, the interests of various stakeholders need to be taken into account, including scientists from different disciplines and local stakeholders with different goals and priorities. Software methods to support participants in these discussions are now widely believed to help make wiser and more sustainable management decisions by more easily weighing up the views of all relevant parties. These parties include land-owners, residents, environmental pressure groups, wildlife biologists and other scientists, governmental bodies and industries. When there is disagreement, people require ways to explore the reasons for the different viewpoints and to seek out areas of consensus which can be built upon. This work uses an AI method based on meta-level representation of argumentation frameworks to explore multiple knowledge bases in which conflicting opinions about environmental change are expressed. A formal meta-language is defined for articulating the relationships between, and arguments for, propositions in knowledge bases independently of their particular object-level representation. A prototype system has been implemented to evaluate the usefulness of this framework and to assess its computational feasibility. The results so far are promising.

This chapter comes from the Scottish pioneer Mandy Haggith at the University of Edinburgh.

Chapter 9   Underground Nuclear Testing
An Intelligent Assistant for Nuclear Test Ban Treaty Verification.
This chapter addresses treaty verification for underground nuclear testbed agreements using a hybrid-AI software agent assistant that classifies and filters seismic data from Norway’s regional seismic array, NORESS.  Verification of a Comprehensive Test Ban Tan (zero testing) has driven the development of enhanced seismic verification technology with lower detection levels and better noise reduction signal extraction algorithms. However each detected event must be analyzed to determine if it contains a clandestine nuclear test. Lowering the detection threshold causes an exponential increase in the number of events detected. The volume of events to be analyzed and classified overwhelms human analysts. SEA was developed in the Treaty Verification Research Group at Lawrence Livermore National Laboratory.  The overall system is hybrid - it contains hardware and many kinds of software, such as advanced signal processing algorithms that work with SEA.  The agent architecture supports a pattern driven application of computationally expensive numerical analysis. Three important aspects of the intelligent software assistant SEA are (1) the user interface permits interactive or human-agent analysis (2) it reduces the workload of the human analyst by filtering and classifying the large volume of continuously arriving data, presenting “interesting” events for human review and explanation of its analysis (3) it emulates the common sense problem solving behavior and explanation capability of the human seismic analyst by using a multi-context Assumption Based Truth Maintenance System.

The work was conducted in the Treaty Verification Program at Lawrence Livermore Laboratory in the USA by Cindy Mason.  Dr. Mason authored the proposal for the first international AI and Environment  workshop while on a joint appointment with Stanford University and NASA Ames Research Center.

Chapter 10  Assembling Satellite Data
The COLLAGE/KHOROS Link: Planning for Image Processing Tasks
This chapter looks at the assembly of satellite data.  It is an overarching and pervasive issue in environmental computer systems.  Challenges include how to represent and partition information in a way that fosters extensibility and flexibility and how to do this across many kinds of satellite data and analysis products that are often changing and growing.  To solve this problem we use a branch of AI known as planning.  AI Planning allows us to automatically generate the necessary sequence of image processing steps for examining satellite remote sensing data. Several obvious issues arise when integrating a variety of data and products for viewing/analysis: low-level connection tasks; representation translation tasks; the need to present different kinds of users with a suitably coherent combined architecture.  To make the system open to future media and data products we are interested in how to represent and partition information in a way that fosters extensibility and flexibility.  We describe our work to do this with two existing systems at NASA called - COLLAGE/KHOROS, which accesses a suite of image processing algorithms that are constantly changing and updating.  Our challenge for the planning system is to provide the assembly and viewing of these data and data products to be useable by a variety of users with different skill levels. These kinds of issues, of course, are common among many software engineering enterprises.

This is a NASA project from the USA.  Pioneers Scott Schmidler and Nick Short from NASA Goddard and Amy Lansky, Mark Friedman, and Lise Getoor, from NASA Ames created the work in this chapter.

Chapter 11   Forest Ecosystem Simulation
KBLIMS For Forested Ecosystem Simulation Management
This chapter addresses forested ecosystem management with hybrid-AI question answering simulation systems.  Answering questions like “what is the effect of clearcutting on  watersheds in the Turkey Lakes region of Ontario Canada?”  involve complex interactions between simulations and specialized tools about climatic, topographic, hydrologic, pedological and ecological processes. The manual process of answering a question is a laborious task where simulations, tools and modeling systems from multiple disciplines are run like batch processing, generating many cumbersome files. The questions also involve the complex interactions between fundamentally different kinds of data from geographic information system(GIS) and ecosystem simulation modeling. For example, geographic information systems typically represent information as points, polygons, lines and layers, whereas simulation systems use system state, mass and energy flux, and the interaction and dynamics of species or individuals. Automating this process relieves the tedium and speeds the process involved in each q/a so that many more queries can be completed. To manage aggregation and integration across these fundamentally different disciplines, data types and systems,  the AI systems use a multi-layered ontology across conceptually different systems with an architecture based on the notion of a query model that executes a set of user-defined queries. The system allows many kinds of queries over many combination and levels of aggregation and scales and includes simulation queries, spatial data queries, deduction queries and an aggregation of these processes.  The system can run on either user-defined or system-defined queries. Typical use of the system is done automatically so the user, e.g. an ecologist, need not explicitly parameterize and run simulation models. System use of the simulation systems are managed by the knowledge base using its meta knowledge about the tools, modelers, etc. which allows for the integration of either tightly-coupled or loosely-coupled systems.   Users interface to the system expressing a simulation experiment by first identifying a set of high level concepts/objects as a spatial query, then specifying some action to be performed on these concepts/objects, such as a combined simulation query and aggregation query.  The AI explanation system is based on the same ontological/concepts.

This chapter is based on work conducted in Canada by pioneers Vincent B. Robinson and D. Scott Mackay while at the University of Toronto.

Chapter 12  Weather Bulletins
SCRIBE: An Interactive System for Composition of Meteorological Forecasts
This chapter describes work by the Canadian Meteorological Center to interactively generate public weather forecasting bulletins from weather matrices and sensors across and people across Canada.  The system, SCRIBE, uses hybrid-AI or ensemble methods to generate plain language public forecast bulletins in French or English from a set of stations or sample points  prepared at a three-hour time resolution over a range of Canada.  Although the system is created for the purpose of automation it is also run in manual mode and all processing can be monitored and modified by human users.  A semantic numerical analysis processes the weather element matrices according to standards of codification.   The resulting content is described with more than 40 precipitation  concepts (rain, rain heavy at times...), including three types of concepts applicable to thunderstorms (risk, possibility, a few) at up to three levels at the same time (ex.: rain and snow possibly mixed with ice pellets) it can also produce two types of concepts applicable to precipitation accumulation (liquid and frozen), six classes of probability of precipitation concepts, 13 sky cover concepts (11 stationary states and two evolving states), 14 classes of wind speed with eight directions, two types of visibility concepts (blowing snow and fog) and ten types of maximum/minimum temperature concepts. By using the standards of codification the AI system provides a simple way to display the content of the weather element matrices for human editing rather than displaying the raw numbers. Once the editing task is complete at the interface level, the modified concept file is quality controlled before being fed to the knowledge base system again to generate the plain language bulletin. The knowledge base system creates a basic sentence structure that can be matched into different structures representing different semantics expressing the same content, following a case base reasoning approach.   The knowledge base system uses approximately 600 rules to generate the standardized ontological weather concepts.  It uses approximately 1000 rules to generate the plain language bulletins. The use of rules supports the ability to explain the steps of the automation process.

The Canadian pioneers for this work conducted at the Canadian Meteorological Center are R. Verret, G. Babin, D. Vigneux, J. Marcoux, J, Boulais, R. Parent, S. Payer, and F. Petrucci.  

Chapter 13  Weather forecasting
Retrieving Structured Spatial Information from Large Databases
This chapter addresses weather forecasting with an intelligent software agent assistant.  The agent acts as a “memory amplifier” for meteorologists to assist in weather forecasting by rapidly locate and analyze similar kinds of past weather.  Much of environmental data, including meteorological, covers a large region of the earth so it is organized as spatially. A challenge is that historical multi media meteorological data includes audio, text, satellite images, laser disks, etc..   The chapter presents the first AI method to intelligently retrieve spatially organized data with a technique known as a case-based reasoning system for the rapid display of historical meteorological data.  Case based analysis allows the comparison of similar instances or ‘cases’ of an example.  This work is distinguished by the large size of its case base, by its need to represent structured spatial information, and by its use of a relational database to store spatial data cases. This  briefly describes some of the technical issues that follow from these design considerations, focusing on the role of the relational database.  The system is called MetVUW Workbench.

This work was conducted in New Zealand by pioneers Eric K. Jones and Aaron Roydhouse while at the University of Wellington.

Chapter 14  Sharing Digital Environmental Resources
Environmental Information Mall  
The chapter addresses sharing environmental tools and data products across government agencies, institutions and other large organizations.  The key to the Environmental Information Mall project is an environmental concepts ontology.  It supports the interoperation of data and analytical tools from a variety of independent sources.  The chapter describes AI tools for the creation and maintenance of the ontology, and then shows how it can be used including but not limited to: information sources advertise their capabilities; mediators combine analytical tools with the data on which they operate to share data products; end users locate relevant information; and intelligent agents or intelligent user interfaces that fuse information from several sources.   

This chapter comes from Texas, USA.  It is based on the work of pioneers Michael Huhns, Munindar P. Singh and Gregory E. Pitts who were working at MCC.

Chapter 15  Biodiversity and ecosystem catalogues
BENE - Biodiversity and Ecosystems Network Environment
The chapter addresses biodiversity and ecosystem cataloging in global collaboratives for biodiversity conservation and ecosystem protection, restoration, and management communities. BENE (Biodiversity and Ecosystems Network Environment) fosters enhanced communications and collaborations through the intelligent sharing of networks of biodiversity and ecosystem data and collections. Current estimates of the diversity of life (plant, animal, microorganisms) on Earth (biodiversity) range beyond the ~1.5 million species described to date up to perhaps as high as -130 million species. The number of entries in such a collective are vast and are created by  wide ranging diverse social, political and economic members from governments, corporations, academia, private foundations, individual citizens and so on.  The data types are also diverse and include  samples or specimens themselves, field notes of scientists and taxonomists, museum repositories, geographic information systems and spatial data, genome data, education and television data, etc.   BENE project a) points users to new networks of biodiversity data collections using intelligent search b) provides web based user access to an integrated network of collectives using search agents that rely on ontological and meta-data. At present, there are nodes in Australia (4), Brazil (the BIN21 Secretariat resides at the Base de Dados Tropical), Costa Rica, Ecuador, Finland (2), Italy, Japan, United Kingdom and the United States (BENE is the only BIN21 node in the U.S.A.).

This chapter represents a joint effort by the USA conducted by pioneering scientist Steve Young at the Smithsonian Institute and the US Environmental Protection Agency and Leland Ellis and Andrew Jackson at Texas A&M University.

Chapter 16   Plant physiology and climate change modeling
Automated Modeling of Complex Biological and Ecological Systems
This chapter concerns plant physiology and climate change modeling, specifically the automatic generation of simulation models that answer prediction questions and explain how climate change may affect plant physiology.   It is particularly useful to predict the effects of global climate changes on plants and animals in specific regions. In general, predicting answers to climate change questions takes vast amounts of knowledge, time, people with special knowledge and is error prone. Automating this prediction process speeds it up allowing consideration of many different scenarios and assumption conditions.  Equally important to the answer for a prediction question is the reason for that answer.  Any system like this must also cough up an explanation.  But the automation tools themselves choke on the vast knowledge during computation.   To answering climate prediction questions we not only need general principles of plant and animal physiology but species interactions and specific data on individual species, climatic events, and geologic formations.  The central issue in automatically answering prediction questions is constructing a model from this wealth of information that captures the important aspects of the scenario and their relationships to the variables of interest.  This avoids the problem of working with a sea of knowledge, much of which is irrelevant to a particular question.  The novel approach taken to solving this problem  makes Siri and Watson look like kindergarten tools.  Spoiler - the AI system automatically generates code on the fly for each prediction question.  The key to this approach lies in building a meta model of the knowledge, causal-relations and tools.  Using the meta knowledge, a predictive question/answering system is coded on the fly based on causal relationship elements needed for each question using only computing simulations and knowledge elements relevant to the question. The causal information is also the basis for the explanation facility.   Consider the general form of a prediction question in plant physiology: “How would decreasing soil moisture affect a plant’s transpiration rate?” A prediction question poses a hypothetical scenario (e.g., a plant whose soil moisture is decreasing) and asks for the resulting behavior of specified variables of interest, (e.g., the plant’s transpiration rate). Using detailed knowledge of plant physiology and other physical systems, the q/a system is generated by parsing the prediction question and determining the relevant elements and factors needed to answer the question.   The authors introduce a modeling program called TRIPEL for answering prediction questions based on causal influences. It defines the modeling task, criteria for distinguishing relevant aspects of the scenario from irrelevant aspects and the algorithm that uses these criteria to automatically construct the simplest adequate model for answering a question.  The chain of causal influences provides the basis for an explanation facility.  Representing the information in multiple levels of abstraction support explanation styles matched to the kind of user (scientist, decision maker, etc.) This system can be generally used on any body of knowledge to automatically generate predictive q/a systems. In biology and ecology, such questions are important for predicting the consequences of natural conditions and management policies as well as for teaching biological and ecological principles.

This chapter represents the work of Texan pioneers Jeff Rickel and Bruce Porter at the University of Texas, USA.

This “Book in a Nutshell” is composed by the editor and contributor, Cindy Mason and is a combination of facts found in the chapters together with her own writing.  Any errors in the representation of the ideas in the chapters are the sole responsibility of the editor.