基于WorldView-3和OP-ELM的矿化蚀变提取方法研究

The spectral differences from visible through the short wave infrared (SWIR) region of the electromagnetic spectrum for alteration minerals are the theoretical basis of the studies for alteration mineral mapping using remote sensing data. Limited by the remote sensing data and the usable methods, alteration minerals with small scale are difficult to identify. This project aims to study the effective method for identifying alteration zones with small scale. The Mingjingou gold deposit and the surrounding area were set as the study area. The WorldView-3 images, field measured spectra and the spectra library were selected for the study. First, based on the spatial resolution and spectral resolution of WorldView-3, the characters of endmember spectra of different alteration types were analyzed. Second, the training samples were selected by overlaying the results of three spectral matching methods. Third, an optimally pruned extreme learning machine was trained and built for distinguishing the scale of alteration zones. Fourth, the OP-ELM medal was improved after the field validation. Finally, the relationship between the changing spatial resolution and the identification ability for the different scale of alteration zones was quantified. The utilized methods and results could provide a theoretical basis and data support for mapping small scale alteration zones and the spatial resolution selection of remote sensing data for remote sensing exploration with different scale.

蚀变矿物在可见光-短波红外波段范围具有光谱特征，是利用卫星遥感图像提取蚀变信息的重要理论依据。由于遥感数据源和可用技术方法的局限，小规模蚀变信息难以有效识别。本项目基于WorldView-3、JPL光谱库及实测光谱数据，在分析不同分辨率下蚀变像元光谱特性变化规律、选取训练样本及确定模型最优参数的基础上，建立优化极限学习机（OP-ELM）模型对各类型蚀变规模进行识别。进一步研究空间分辨率变化对不同规模蚀变提取精度的影响，量化各空间分辨率下可辨识蚀变的最小规模，提出适用于小规模蚀变信息提取的有效方法。研究结果可为小规模遥感蚀变提取、不同比例尺遥感找矿工作中数据分辨率选择提供理论基础与数据支持。

为研究遥感影像像元空间分辨率变化对小规模蚀变信息提取能力的影响，本课题在数据预处理的基础上，将WorldView-3数据可见光-短波红外波段数据像元重采样至5米、10米、15米、20米和30米5种分辨率模式。像元光谱分析表明面状蚀变由于面积较大，空间分辨率变化对像元光谱特征影响较小，吸收特征得以保留。较宽脉状蚀变带（宽度8-25米）随着空间分辨率的降低，像元光谱受混合像元影响越来越强，导致像元光谱特征吸收越来越浅，30米空间分辨率时吸收特征最弱，但仍可以被有效识别。较窄脉状蚀变带（宽度2-5米）随着空间分辨率的降低，混合像元影响越来越强，导致像元光谱的吸收特征越来越浅，30米空间分辨率时吸收特征最弱，基本无法有效识别。. 根据已知地质图、野外验证点并结合光谱匹配滤波法选取分类训练样本。建立了OP-ELM模型，通过试验确定了隐藏层神经元和训练样本的数量，获得研究区岩性、蚀变带分类图。根据野外定位的32个检查点和3个未检查的矿点，对分类精度开展评价，结果表明分类结果总体精度较高，达到了94.3%，仅2个位置与验证点不一致。证实了OP-ELM方法开展Worldview-3数据蚀变分类具有较好的精度。. 利用光谱匹配滤波法分别处理不同分辨率短波红外数据，获得了不同空间分辨率影像白云母矿物提取结果，宽度约8-25米的线状蚀变带在5-20米空间分辨率数据提取结果中均线状特征明显，在30米空间分辨率结果中仅断续出现。宽约2-5米的蚀变带在5-15米空间分辨率图像中均线状特征明显，20米、30米空间分辨率结果中仅零星像元被突出，难于识别。对于面状蚀变带5种分辨率均可较好提取，空间分辨率变化对提取结果影响不明显。因此，考虑后续在脉状目标提取方面的大量应用，建议后续多光谱、高光谱传感器设置的空间分辨率优于20米，方可有效提取各类形态规模的蚀变带。