摘要:
影像解译中对土地利用单分类的关注成为遥感研究的热点问题。最大熵模型(MaxEnt)被评价为最有潜力的单分类算法,被广泛应用于土地利用的单分类研究。然而,单分类算法(包括MaxEnt)是否能够进行土地利用多分类尚不明晰。为了解决该问题,文章建立了利用MaxEnt进行遥感土地利用多分类的技术流程,并将该流程应用在云岩河流域的土地利用多分类研究中。使用总体分类精度、Kappa系数、灵敏度以及特异度评估MaxEnt的总体分类效果以及在各个地类上的预测表现; 同时使用Kappa值评估MaxEnt与随机森林(randem forest,RF)、最大似然法(maximum likelihood classification,MLC)、支持向量机(support vector machine,SVM)在土地利用预测上的一致性表现。结果表明: ①MaxEnt的分类表现最好,总体分类精度为84%,Kappa系数为0.8; ②MaxEnt在各个地类上没有最差的表现,甚至在某些地类上达到了最优的表现; ③MaxEnt与RF和SVM的分类一致性较高,这3种算法预测的土地利用之间一致性评估Kappa值均超过了0.6; ④MLC与其他3种分类算法预测土地利用的差异较大,Kappa值小于0.4,说明MLC不适合该地区的土地利用解译。文章建立的技术流程仅仅依赖于土地利用发生概率,而不依赖于阈值选择,从而使得以MaxEnt为代表的单分类算法在遥感土地多分类应用中能够发挥巨大潜力。对于大范围的土地利用解译,加入并行计算将有利于提高利用MaxEnt解决多分类问题的时间效率。
Abstract:
The one-class classification (OCC) of land use in image interpretation is a hot research topic of remote sensing. Many novel algorithms of OCC were introduced and developed. The maximum entropy model (MaxEnt)-the most promising OCC algorithm as evaluated-is widely used in the OCC study of land use. However, it is unclear about the applicability of these algorithms (including MaxEnt) in multi-class classification (MCC) of land use. Thus, this study established a procedure for MaxEnt-based land-use MCC in remote sensing image interpretation and applied the procedure to the land-use MCC of the Yunyan River basin. The overall classification effect of MaxEnt and the performance of MaxEnt in the prediction of various land were evaluated using overall classification accuracy, Kappa coefficient, sensitivity, and specificity. Moreover, the Kappa coefficient was also used to evaluate the consistency between MaxEnt and random forest (RF), maximum likelihood classification (MLC), and support vector machine (SVM) in the prediction of land use maps. The results are as follows: ① MaxEnt showed the best classification effect, with overall classification accuracy of 84% and a Kappa coefficient of 0.8; ② MaxEnt showed no worst performance in any land type, and even performed the best in some land types; ③ MaxEnt showed high classification consistency with RF and SVM, and the consistency evaluation of the land use maps obtained using the three algorithms yielded Kappa coefficients of greater than 0.6; ④ Compared with the other the three algorithms, MLC yielded a significantly different land use map, with a Kappa coefficient of less than 0.4. This result indicates that MLC is not applicable to the interpretation of land use of the study area. The procedure established in this study only depends on the occurrence probability of land use rather than the threshold selected. As a result, the OCC algorithms represented by MaxEnt have great potential for application to the land-use MCC in remote sensing image interpretation. In addition, the introduction of parallel computing into large-scale land use interpretation will help improve the efficiency of solving MCC problems using MaxEnt.
下载: