Assessment of Landuse and Land Cover Change Using Remote Sensing and Artificial Neural Networks

SBR-9513889由该合同资助的研究将测试人工神经网络（ANN）在检测遥感图像变化中的效用。探测和监测地球表面条件变化的能力对于了解人类对环境的影响和评估发展的可持续性至关重要。地面测量受到后勤限制的严重限制，特别是在全球变化的研究和评估方面。因此，遥感提供了频繁评估大面积地表条件及其变化的可能性。目前，通过遥感数据进行变化检测的最大成功是针对剧烈变化的情况。这项由该奖项资助的研究的目标是测试人工神经网络在评估不太剧烈的变化方面的效用。人工神经网络将在变化检测方面取得重大进展的期望是基于两个因素：人工神经网络已被证明比传统的基于统计的遥感图像分类方法更有效，并且研究人员在以这种方式使用人工神经网络的初步努力中取得了一些成功。测试需要几个组成部分：探索可选择的人工神经网络架构，解释人工神经网络的内部结构以分析遥感数据中的变化信号，相对于传统的变化检测方法评估人工神经网络的性能，开发解释人工神经网络输出信号以量化变化的措施，以及估计训练后的人工神经网络在训练域外的鲁棒性。调查人员将使用现有的数据集，其中包括所有这些测试的遥感和地面测量数据，从而以最小的成本取得快速进展。该研究具有理论和应用意义。它将导致对陆地表面变化产生的遥感图像中包含的光谱和时间信号有更大的概念理解。它应该有助于确定哪种分析方法更适合于在不同的环境中测量变化。该研究将有助于提高对神经网络在数据分析框架中的应用的理解，从而增强神经网络在地理研究中的应用。

SBR-9513889 The research funded by this award will test the utility of artificial neural networks (ANN) in detecting changes in remotely sensed images. The ability to detect and monitor changes in conditions at the Earth's surface is fundamental to an understanding of human impacts on the environment and to the assessment of the sustainability of development. Ground based measurements are severely limited by logistical constraints, particularly in the research and assessment of global change. Therefore, remote sensing provides the potential for frequent assessment of surface conditions and their change over large areas. Currently, the greatest success in change detection via remotely sensed data is for situations of dramatic change. The goal of the research funded by this award is to test the utility of ANN in assessing less dramatic change. The expectation that ANN will allow significant improvements in change detection is based on two factors: ANN have proven more effective than conventional statistics-based methods for classification of remote sensing imagery, and the investigators have had some success in a preliminary effort to use ANN in this way. The testing entails several components: exploring alternative ANN architectures, interpreting the internal structure of the ANN to analyze the changes signal in the remotely sensed data, evaluating the performance of the ANN relative to conventional methods for change detection, developing measures to interpret ANN output signals for quantifying change, and estimating the robustness of trained ANN outside of the training domain. The investigators will use existing data sets containing remotely sensed and ground measurements for all these tests, allowing rapid progress at minimal cost. The research has both theoretical and applied implications. It will lead to a greater conceptual understanding of the spectral and temporal signals contained in remotely sensed images resulting from land surface change. It should help determine what analytic methods are better suited to measuring change in different contexts. The research will lead to an improved understanding of the use if neural networks in a data-analytic framework, and thus enhance the appropriate is of ANN in geographic research.