Conference Agenda

Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).

 
Session Overview
Session
320R: Artificial intelligence and machine learning for land use modelling
Time:
Wednesday, 24/Apr/2019:
2:00pm - 3:30pm

Session Chair: Bumsuk Seo
Session Chair: Calum Brown
Location: MB-114
Main Building, 1st floor, west wing, 78 seats
Session Topics:
How do we support transformation?

Session Abstract

This session will focus on new applications using artificial intelligence (AI) and machine learning (ML) in land use modeling. Identifying and reducing uncertainties in future land use projections are critical to integrated assessments of climate and social change scenarios. However, uncertainties related to model design and fit-to-data remain substantial. AI has great potential to improve the predictive performance of land use models following breakthroughs in satellite-based land use classification, social media data analysis, and heterogeneous data mining. However, the application of these techniques in land use modeling is still limited.

On the other hand, we need to think further how LSS and data science can learn from each other. The LSS community has a long history of utilising extremely heterogeneous data and highly-abstracted concepts for modelling complex socio-ecosystems. We are therefore in a good position to share insights and advice with the AI / Big Data communities.

This session aims to bring together land use and land cover researchers and to exchange ideas and experiences on all aspects related to the application of AI and ML methods. Of particular interest is mining of heterogeneous ‘big data’ sources, such as social media data, camera trap and drone video data, large-scale statistics, and grey literature, all of which can provide information about land use and decision-making process. AI-based methods are used to process multimedia data (e.g., drone video) data for mapping land use and land cover in natural and urban areas. New ML-methods such as deep reinforcement learning can potentially provide a basis for robust calibration of complex land use models by exploring large parameter spaces efficiently. Overall, these established data sources in combination with new AI and ML technologies can open up new perspectives in land use modelling. It is also crucial in the session to have fruitful discussions about how to synthesize the lessons and insights for giving advice back to the AI / Big Data communities, so the learning goes both ways.

The session intends to cover a wide range of ML topics related to land use models, including methodologies, technologies, empirical and experimental studies. We welcome submissions of AI and ML papers providing new insights into the future of land use models.

Session Organizers: Bumsuk Seo, Calum Brown, and Donggul Woo


Show help for 'Increase or decrease the abstract text size'
Presentations
Full talk
ID: 449 / 320R: 1
320R Artificial intelligence and machine learning for land use modelling
Keywords: Land Use Modelling, Artificial Intelligence, Machine Learning, Land System Modelling, Neural Networks

How could land use models benefit from recent developments in Artificial Intelligence: promises and challenges for large-scale land use modelling

Bumsuk Seo1, Calum Brown1, Mark Rounsevell1,2

1Karlsruhe Institute of Technology, Germany; 2University of Edinburgh, UK

Land use models are often complex to construct and difficult to validate especially at large spatial scales. This is especially challenging for agent-based and similar behavioural modelling approaches that provide flexible descriptions of land use decision-making. New Artificial Intelligence (AI) and Machine Learning (ML) techniques are emerging across various scientific disciplines. AI, especially recent development in deep learning has the potential to improve complex adaptive models by harnessing atypical big-data sources. This presentation aims to raise awareness of new methodological developments and to highlight current achievements as well as gaps that need to be resolved. A general introduction to the state-of-the-art in AI approaches relevant to land use modelling will also be presented. Image recognition techniques can transfer a vast amount of image and video data to land use models. We introduce a European-scale case study using social media data with Convolutional Neural Networks. The approach uses reinforcement learning (e.g., Deep Q-learning) to enable an AI agent to play `land use' games that find the best land management strategy. In this way, we see a large potential for calibrating high-dimensional land use models. A further approach is based on augmenting missing input data using deep Generative models that aims to simulate future land use patterns. Finally, we propose a land use modelling framework that combines agent-based concepts, robust social theories, and AI techniques. This framework is designed to assess land use policy actions in the rapidly changing global environment. We also aspire to trigger a fruitful discussion on what we the land use modelling community can offer to AI and ML communities. This would build on the lessons learned from long-term experience of dealing with complex human-environment systems. This presentation strives to raise awareness of new methodological advances as well as helping to improve future land use projections



Full talk
ID: 448 / 320R: 2
320R Artificial intelligence and machine learning for land use modelling
Keywords: Habitat Modelling, Conservation planning, Artificial Intelligence, Network Analysis, Yellow-throated marten

Habitat modelling and conservation area planning using camera-trap photos, computer vision, and network analysis: a case study of Yellow-throated marten in South Korea

Donggul Woo1, Bumsuk Seo2, Tae-Young Choi1

1National Institute of Ecology, Republic of Korea; 2Karlsruhe Institute of Technology, Germany

Toward improved wildlife and habitat conservation, a new data-driven conservation area management framework is presented in this talk. One of the major issues in conservation area management is to map wildlife habitats for optimal planning. Recent development in Artificial Intelligence (AI) and ML (Machine Learning), especially in image recognition techniques has the potential to derive such distributions using formal camera trap photos as well as casual social media photos. We present a study which explored the potential of these new AI and ML techniques in identifying animal species, thereby estimate their habitat distributions in forest patches. Focused on Yellow-throated marten (Martes Flavigula), we developed a protected area recommendation system based on these habitat modelling. For eight years (2010-2017) in 12 wildlife hotspots in South Korea, we collected wildlife occurrence data using camera traps (HC 600, Reconyx) and identified the animal species using Convolutional Neural Networks. The resulting occurrence data was used to estimate the temporal occurrence pattern of the target species. In parallel to that, network properties and landscape indices of the camera trap patches were calculated. Using land use, forest canopy, and digital elevation model, we created a forest patch network of the 10 km buffer around the camera patches and estimated habitat quality of the camera patches. Potential distribution of the target species was estimated by combining the camera-trap based occurrence model and the habitat quality the entire forest patches. We confirmed that considerable improvement in habitat modelling by using camera photos but it would be needed to use more citizen science and other various data sources. By combining camera trap data, machine learning techniques, and network analysis, the proposed approach is to establish essential conservation areas and help designing conservation plans.



Full talk
ID: 762 / 320R: 3
320R Artificial intelligence and machine learning for land use modelling
Keywords: Land use modeling, machine learning, regression

The land use modeling by iterative parametrization with machine learning approaches

Yu-Pin Lin, Wan-Yu Lien, Yen-Sen Lu

National Taiwan University, Taiwan

Land use change models have been developed and widely used to simulate spatial and temporal patterns of land conversion and to understand the causes and consequences of such changes, particularly in land use planning and conservation initiatives. However, the complexity of land system leads to the poor performance of some land use change models. The notable land use change model is the Conversion of Land Use and its Effects (CLUE-s) model which uses logistic regression (LR) techniques to estimate suitability maps. However, quantifying all the potential interactions between the different drivers of land use in a logistic regression model is difficult. This study uses alternative techniques, including Artificial Neural Networks (ANNs), Random Forest (RF), and Support Vector Machine (SVMs), to estimate suitability maps for calibrating land use change model based on land use patterns and develops an iterative procedure for all of the above techniques to simulate land-use patterns. To avoid overfitting problems and improve model fidelity, each land use suitability modeling technique is compared and contrasted via Area Under the Curve (AUC). In this study, the five counties, including Miaoli County, Taichung City, Nantou County, Yunlin County and Changhua County, with a total area of 10506.8Km2 in central Taiwan, are selected as study areas to evaluate the predictive power of the three techniques. The results indicate that the model based on RF exhibited a higher accuracy than the models based on ANN and SVM. The results also point out the benefits of comparing multiple suitability models based on AUC when selecting suitability models for land-use change models.



Flash talk
ID: 477 / 320R: 4
320R Artificial intelligence and machine learning for land use modelling
Keywords: Brazil, cattle, deforestation, Machine Learning, Big Data

Using Machine Learning to improve understanding of land use in Brazil’s cattle sector

Lisa L. Rausch, Matthew Christie, Jacob Munger, AnHai Doan, Amintas Brandao, Philip Martinkus, Holly K. Gibbs

University of Wisconsin - Madison, United States of America

Brazil’s cattle sector remains responsible for the majority of Amazon deforestation in spite of zero-deforestation agreements that, in principle, cover most of the sector. However, after a decade of partial implementation, under which slaughterhouses only monitor deforestation on their direct suppliers and ignore other properties involved in producing the cattle, evidence is accumulating that full implementation will be necessary to maximize their impact and reduce leakage. Failure to fully implement has been, in part, a data problem; companies lack data that identifies their indirect suppliers and allows them to assess their deforestation. Likewise, researchers and policy makers have been unable to fully quantify the scope of the problem of incomplete implementation due to the same data gap.

We found that datasets that detail farm characteristics, and that detail animal movements between farms, are available, but assembling them to create a full picture of the supply chain requires scripted queries of websites at multiple public agencies and addressing numerous inconsistencies within and between datasets. To solve this critical challenge, we developed a data lake to store an assortment of raw data, including both spatial and tabular data; a data warehouse to normalize and store cleaned and structured data; and an input-output logical processing pipeline to merge data from the warehouse into an abstracted database layer for real-time analysis, resulting in a cloud-based research platform that integrates millions of animal transactions between properties and slaughterhouses with spatial data on hundreds of thousands of properties in the Amazon. We then employed a ML-based entity-matching system to improve the normalization process and, therefore, the accuracy of outputs. I will explore how our LSS/Data Science collaboration advances research on land use decisions related to Brazil’s cattle sector and in Data Science, as well as improves policy implementation.



Full talk
ID: 829 / 320R: 5
206R Relevance of long-term land-use change for sustainable land management
Keywords: agricultural landscape, soil erosion, sediment, water quality

The effect of land use on surface water pollution from non-point sources

Tomas Dostal, Miroslav Bauer, Josef Krasa, Barbora Jachymova

CTU Prague, Czech Republic

Land use related to intensive agriculture production is the most important non-point pollution source for surface water. The pollution can be mainly characterized by erosion and sediment transport processes and can be identified as suspended solids and erosion Phosphorus occurrence in streams, rivers and water reservoirs, with all related consequences.

Catchment with size of 35000 km2 has been analyzed from point of view of input of erosion sediment and erosion Phosphorus from single fields and pollution has been tracked from single field until its deposition within water reservoir. The methods used were mathematical model WATEM/SEDEM and advanced methods of GIS application using hydrological network structuring.

The sediment and Phosphorus load has been balanced within more than 170.000 critical points at water courses and water reservoirs. Points has been classified into 5 categories, according its pollution load.

The effect of recent land-use combined with landscape morphology on pollution load has been studied and optimum soil erosion control measures has been designed at selected most critical subcatchments. Different measures have been assessed and their control effect has been modeled. There has been identified and documented, that diverting technical soil erosion control measures have low or even negative impact on sediment and especially erosion Phosphorus delivery into water bodies. Therefore, soft control measures, based mainly on spatial grassing of arable land and application of conservation tillage practices or nature based buffer zones are necessary to limit agricultural pollution of water courses. Different scenarios of conversion of arable land into grassland has been examined and optimization procedure, related to morphology and required effect on sediment and Phosphorus load reduction has been documented and presented.

The research has been supported by projects No. Mobility7AMB18DE006; SHui 773903; COST LTC18030; SGS17/173/OHK1/3T/11 and QK1720289.



Full talk
ID: 569 / 320R: 6
107R Assessing, modelling, and analysing land use and land management impacts on the Earth system
Keywords: Remote sensing, Landscape ecology, Conservation, Machine learning, Roads

Automated rural road detection and classification to support analysis of the impacts of land use and management

Sean Patrick Kearney1, Nicholas C. Coops1, Gord B. Stenhouse2, Simran Sethi1

1University of British Columbia, Canada; 2fRI Research, Canada

Road access is a strong predictor of both human well-being and land degradation, and detailed spatial road network data is essential for strategic planning and assessing environmental impacts associated with land management. Globally, road length now measures in the tens of millions of kilometers, and it is estimated that about 65% of all roads are unpaved, with variable levels of accessibility. Currently, we lack accurate and up-to-date spatial datasets of road networks and their accessibility, especially in rural areas with roads that are highly dynamic over space and time. We present a method to automatically extract roads from satellite imagery and classify them in a way that reflects accessibility and traffic. We used a combination of machine learning and crowd-sourced ground data to detect roads, evaluate surface conditions and estimate traffic based on population sources (e.g., urban settlements) and sinks (e.g., industrial and recreation destinations) for an intensively managed rural landscape in Alberta, Canada. The resulting road network map was then analyzed to better understand how industrial and recreational road use is affecting a threatened population of grizzly bears. In Alberta, this work has implications for refining wildlife conservation policies related to road construction and access, which currently rely on out-dated and coarsely classified road types. More broadly, the ability to create up-to-date and accurately classified road network maps from satellite imagery and crowd-sourced datasets would support more nuanced analyses of land management impacts and their relationship to road access.



 
Contact and Legal Notice · Contact Address: osm2019@glp.earth
Privacy Statement · Conference: OSM 2019
Conference Software - ConfTool Pro 2.6.127+TC+CC
© 2001 - 2019 by Dr. H. Weinreich, Hamburg, Germany