基于时间序列遥感信息和深度学习的山区撂荒地识别模型研究

The abandoned farmlands are usually characterized by small area, scattered distribution and spectral heterogeneity in mountainous areas due to the topographic variability. It is difficult to obtain their spatial and temporal dynamics using the remote sensing technology, which directly affects the accuracy of the dynamic change of the farmlands. This project builds an optimized model for the detection of abandoned farmlands in mountainous areas, based on the deep learning algorithm (Convolutional Neural Network, CNN) and time-series 30 m-resolution remotely sensed observations. To solve the problem concerning the lack of training samples, the transfer learning mechanism is introduced during the modeling process. The best strategy for using the time-series data sets and the optimal time interval of the time-series data sets are selected by a series of comparative experiments. The contributions of the deep learning algorithm and time-series observations for improving the accuracy of the detection of the abandoned farmlands are qualitatively and quantitatively analyzed. The key scientific problem expected to be solved in this project is whether the deep learning algorithm extracts the deep-level features used for identifying the abandoned farmlands from time-series remote sensing data. The implementation of this project will provide an alternative solution for land cover classification and target recognition by using the deep learning algorithm and time-series observations. It will not only offer methods for acquiring the spatio-temporal dynamic information of the abandoned farmlands over a large region, but provide data and scientific support for the formulation of China's food security strategy and the construction of the ecological civilization in our country.

受起伏地形的影响，山区撂荒地呈现面积小、分散分布、光谱异质性大等特征，导致基于遥感手段难以获取高精度的山区撂荒地时空动态信息，也直接影响到区域尺度耕地动态变化信息的准确性。本项目以30m分辨率的时间序列遥感信息为输入数据，引入迁移学习机制，采用深度学习算法CNN构建山区撂荒地识别模型，从而获取典型山区高时空分辨率撂荒地动态变化信息。在模型构建中，通过设计对比实验，优选时间序列遥感信息的使用策略，并权衡最佳时间间隔，从而确定最优识别模型。本项目将解决深度学习能否从时间序列遥感信息中挖掘出有利于撂荒地识别的深层次特征这一关键科学问题，也将定量评估时间序列遥感信息和深度学习算法在提高山区撂荒地识别精度中的贡献。项目的实施将为深度学习算法处理基于时间序列遥感信息的分类与目标识别等问题提供解决方案，也为获取大区域山区撂荒地的时空动态信息提供方法支撑，并服务于我国的粮食安全战略的制订和生态文明建设。

在城市化进程不断加速、国家重大生态保护工程大范围实施的背景下，山区耕地撂荒成为一个普遍的现象。受山区起伏地形和多变气候的影响，山区撂荒地地块呈现面积小、分散分布等特征，降低了基于遥感技术自动提取山区撂荒地信息的精度，也成为制约区域尺度耕地动态变化信息提取质量进一步提高的瓶颈。本项目基于自适应时空数据融合模型（STARFM）算法获得了时间序列高空间分辨率遥感数据集，依托深度学习算法（1-D CNN、Transformer和3D-CNN）和常规机器学习算法（随机森林）构建了不同的山区撂荒地识别模型，并分析了物候信息、影像光谱、地形等信息在提高山区撂荒地自动识别精度中的贡献。研究结果表明：（1）STARFM能够实现高空间分辨率影像和高时间分辨率影像的融合，进而挖掘出有利于识别撂荒地的各类物候信息；（2）基于随机森林模型构建的山区撂荒地识别精度为88.91%，Kappa系数为0.87，优于深度学习算法（Transformer）构建的识别模型的精度（79.73%）和Kappa系数（0.76）；（3）不同的特征变量对于撂荒地识别的精度影响各不相同，整体来看，各输入特征的重要性排序为：物候特征 > 影像光谱 > 地形信息 > 各类指数。项目的实施将为深度学习算法处理基于时间序列遥感信息的分类与目标识别等问题提供解决方案，也为获取大区域山区撂荒地的时空动态信息提供方法支撑，并服务于我国的粮食安全战略的制订和生态文明建设。