基于深度学习算法的森林扰动类型遥感分类方法研究

Forest disturbances threaten the ecosystem services provided by forests. Attributing forest disturbance types facilitates long-term forest monitoring and helps forest management policy making, in the face of climate change. However, traditional remote sensing classification algorithms didn’t achieve high accuracy and had difficulty in separating various disturbance types. Deep learning algorithms have been proven to out-perform traditional algorithms in many remote sensing applications and thus are more promising in classifying forest disturbance types. However, large amounts of labeled training samples are essential for deep learning algorithms. It is costly and challenging to acquire such disturbance training samples, and the performance of different deep learning algorithms (i.e. RNN, CNN and combine CNN with machine learning algorithms) has not been tested in the field of forest disturbance attribution. This project plans to leverage state-of-art deep learning algorithms to classify the major forest change agents (i.e. forest fire, harvest, insect outbreak, hurricane and urbanization etc.), and endeavors to reduce the limit of training sample quantity on the performance of deep learning algorithms and quantitatively evaluate the accuracies of different deep learning algorithms in this field of forest disturbance type attribution. Results from this project will provide more accurate and effective technical approach for attributing forest disturbance types, and provide support data for national and regional sustainable development planning.

森林扰动威胁森林的各项生态服务功能。对森林扰动类型的遥感识别有利于在大的时空尺度上对森林进行监测，和制定气候变化背景下的森林管理政策。传统遥感分类方法精度有限，难以有效区分多种森林扰动类型。相比之下，深度学习算法在遥感信息提取方面优势明显，更适合监测区分森林扰动类型。但由于深度学习需要大量标记样本，而获取森林扰动标记样本难度大、成本高，且不同深度学习算法（如RNN，CNN和CNN结合机器学习等）对于森林扰动类型分类的表现仍不确定，因而限制了深度学习算法在此领域的应用。本项目拟开展基于深度学习算法的多种森林扰动类型（如火灾、砍伐、病虫害、风灾和城市化等）的遥感监测分类研究，减小标记样本数量对深度学习算法在森林扰动类型遥感监测分类中的限制，并对不同深度学习算法在此领域的表现进行定量评估。研究成果将为森林扰动类型遥感分类提供更加精确、高效的技术途径，为国家和地区可持续发展规划提供基础数据。

森林干扰威胁森林的各项生态服务功能。对森林干扰类型的遥感识别有助于在大的时空尺度上对森林进行监测，和制定气候变化背景下的森林管理政策。传统遥感分类方法精度有限，难以有效区分多种森林干扰类型。相比之下，深度学习算法在遥感信息提取方面优势明显，更适合监测区分森林干扰类型。但由于深度学习需要大量标记样本，而获取森林干扰标记样本难度大、成本高，且不同深度学习算法对于森林干扰类型分类的表现仍不确定，因此限制了深度学习算法在此领域的应用。.按照既定研究计划，本项目圆满地完成预期目标，并在研究深度及广度上有了进一步突破。主要完成：（1）建立全球典型森林干扰区域（中国东北林区、美国加州和巴西亚马逊）主要干扰类型（如林火和采伐）样本数据库；（2）测试自动生成森林干扰标记样本的方法，并评估其对提升森林干扰分类精度的贡献；（3）评估不同深度学习方法对基于遥感时序数据的森林干扰分类的效果；（4）测试基于深度学习方法的森林干扰分类模型的时空迁移性。在项目资助下，发表高水平学术论文5篇（均发表在中科院一区期刊），主持人获批湖北省楚天学子、武汉市黄鹤英才优秀青年人才和华中师范大学桂子青年学者。