基于分层非监督分类的油菜面积识别研究

Rape Growing Area Identification Based on Layering Unsupervised Classification

中国油菜生产表现为破碎化种植的特点,常规地面调查方法费时费力代表性差,遥感方法逐渐成为了油菜面积监测的理想途径。目视解译、监督分类等遥感分类方法人机交互量大,受监测者主观影响较大。针对上述问题,笔者提出一种基于先分层后进行非监督分类的油菜监测新方法,利用从江县2013年的GF影像对该方法进行了应用和精度评价,本次应用中使用了等距分层和自然分层两种分层方法。结果表明:相较直接非监督分类结果,先分层后非监督分类方法显著提高了总体精度。基于等距分层和自然分层方法总体精度从79.16%提升到了84.44%和85.17%。直接非监督分类中精度仅为72.97%的用户精度在等距分层和自然分层处理后,精度分别提升为81.05%和86.12%,大大降低了直接非监督分类中非油菜区被错判为油菜的现象。等距分层和自然分层方法的总体精度分别为84.44%和85.17%,Kappa系数分别为0.69和0.70。精度上两种分层方法间无显著的差异,自然分层方法的用户精度、制图精度和总体精度都保持在较高的水平,具有更高的可靠性。文中提出的新方法具有人工干预少,精度高的特点,在基于大批量影像的面积监测方面具有较大的应用潜力。

The fragmentation of rape planting in China significantly increases the difficulty of monitoring its growing area, as the regular ground survey takes much time and energy with low representation, remote sensing has become the most ideal method for monitoring rape growing area. The common remote sensing monitoring methods including visual interpretation and supervised classification are comparatively more influenced by the subjectivity of the monitors, due to the high degree of human-computer interaction. Aiming at the above problems, this article proposed a new monitoring method on the basis of unsupervised classification after layering. The GF images of 2013 in Congjiang County were adopted for application and precision evaluation by equidistance layering and natural layering. The results showed that the unsupervised classification after layering was higher in the overall precision than that of the direct unsupervised classification. The overall precisions on the basis of equidistance layering and natural layering increased from 79.16% to 84.44% and 85.17%, respectively. The user precision in the direct unsupervised classification was only 72.97%. After equidistance layering and natural layering, the precision rate was raised up to 81.05% and 86.12% respectively, leading to a great reduction of misjudging the non-rape area as the rape area. The overall precisions of equidistance layering and natural layering were 84.44% and 85.17%, with the Kappa coefficient of 0.69 and 0.70. There was no significant difference in the precision between the two layering methods, hut the natural layering had a comparatively higher degree of user precision, mapping precision and overall precision with higher reliability. The new methods proposed in this article have little manual intervention and high precision, and a large potential to be applied in area monitoring on the basis of mass images.