本文为加拿大萨斯喀彻温大学（作者：Dennis Correa Duro）的博士论文，共191页。
在未来几十年中，随着世界人口的持续增长，以及对西方饮食需求的增加，目前大量的农业生产地区将面临越来越大的压力。再加上气候变化的潜在后果，以及当前能源密集型农业生产方式所涉及的成本不断增加，实现环境和社会经济可持续性的目标将变得越来越具有挑战性。至少，要实现这些目标，就需要对时间和空间的变化率有更深入的了解，以便正确评估目前的需求如何影响子孙后代的需要。
由于农业是现代社会的一个基本组成部分，也是地球上人类引起的景观变化最普遍的形式，因此，测绘和跟踪这些环境的变化是朝着实现可持续性目标迈出的关键的第一步。鉴于对与农业发展相关的一致和及时信息的需求日益增加，本论文提出了地理信息学领域的若干进展，并在遥感和空间分析领域做出了具体贡献：
第一，利用基于像素和基于对象的两种图像分析方法，评估了用于遥感图像分类的几种有监督机器学习算法的相对优势。
其次，基于随机森林分类器的特征选择过程应用于大数据集，以减少分类模型使用的基于对象的预测变量总数，而不牺牲总体分类精度。
第三，引入了一种基于对象的混合变化检测方法，该方法能够处理不同的图像源，生成每类的变化阈值，并最小化地图更新错误。
第四，对粗比例尺农业普查数据进行了空间分解，以明确的空间方式显示萨斯喀彻温省西南部9000平方公里流域的农业发展指标，时间跨度为几十年。所采用的方法组合代表了一个适合支持农业环境可持续发展的总体分析框架。
Over the coming decades it is expected that the vast amounts of area currently in agricultural production will face growing pressure to intensify as world populations continue to grow, and the demand for a more Western‐based diet increases. Coupled with the potential consequences of climate change, and the increasing costs involved with current energy‐ intensive agricultural production methods, meeting goals of environmental and socioeconomic sustainability will become ever more challenging. At a minimum, meeting such goals will require a greater understanding of rates of change, both over time and space, to properly assess how present demand may affect the needs of future generations. As agriculture represents a fundamental component of modern society, and the most ubiquitous form of human induced landscape change on the planet, it follows that mapping and tracking changes in such environments represents a crucial first step towards meeting the goal of sustainability. In anticipation of the mounting need for consistent and timely information related to agricultural development, this thesis proposes several advances in the field of geomatics, with specific contributions in the areas of remote sensing and spatial analysis: First, the relative strengths of several supervised machine learning algorithms used to classify remotely sensed imagery were assessed using two image analysis approaches: pixel‐based and object‐based. Second, a feature selection process, based on a Random Forest classifier, was applied to a large data set to reduce the overall number of object‐based predictor variables used by a classification model without sacrificing overall classification accuracy. Third, a hybrid object‐ based change detection method was introduced with the ability to handle disparate image sources, generate per‐class change thresholds, and minimize map updating errors. Fourth, a spatial disaggregation procedure was performed on coarse scale agricultural census data to render an indicator of agricultural development in a spatially explicit manner across a 9,000 km2 watershed in southwest Saskatchewan for three time periods spanning several decades. The combination of methodologies introduced represents an overall analytical framework suitable for supporting the sustainable development of agricultural environments.
- 农业环境中土地覆盖与土地利用的遥感与地理信息系统方法
- 基于像素和基于对象的图像分析与机器学习算法在SPOT-5 HRG图像农业景观分类中的比较
- 基于对象的多尺度图像分析与基于随机森林的多传感器地球观测图像特征选择
- 一种混合的基于对象的变化检测方法（用于历史景观重建中的多传感器数据集）
- 利用遥感观测和空间分解解释加拿大萨斯喀彻温省西南部大流域土地覆盖和农业用地强度的变化（1976-2005）
更多精彩文章请关注公众号：