遥感时序光谱重构的耕地信息提取方法

Extraction method of cultivated land information based on remote sensing time series spectral reconstruction

针对传统耕地提取方法人工干预多、提取速度慢、成本高,不适用于大规模耕地提取等问题,选取时间序列Sentinel-2数据,基于时序光谱特征的耕地自动识别方法对玛纳斯县典型耕地区域进行耕地提取,并比较了不同样本特征对耕地提取精度的影响。结果表明:采用该方法总体分类精度、Kappa系数、耕地类型的用户精度分别达到了95.37%、0.9423和97.04%,比采用时间序列NDVI和单时期影像样本特征总体精度分别提高4.02%和5.69%。研究结果为进一步利用中高分辨率遥感数据和深度学习方法对耕地进行信息提取和典型地物分类提供了新思路。

Aiming at the problems of traditional cultivated land extraction methods,such as manual intervention,slow extraction speed and high cost,they are not suitable for large-scale cultivated land extraction.In this paper,we extracted the typical cultivated land area from the Manas County and compared the influence of the different sample characteristics on the extraction precision of cultivated land,basing on the automatic recognition method of the time series spectral features,using of the time series Sentinel-2 data.The result shows that the overall classification accuracy,Kappa coefficient and the user accuracy of cultivated land type reached 95.37%,0.9423 and 97.04%.The overall accuracy is improved by 4.02%and 5.69%,compared with the use of time series NDVI and single-period image samples.The research results provide a new idea for information extraction and typical land classification of cultivated land by using the medium and high resolution remote sensing data and the deep learning methods.