许淇　李启亮　Mathilde De Vroey　张吴平　范锦龙

摘要：作物类型遥感识别是农业遥感的重要组成部分，为获取作物种植面积、长势信息并进行产量估算提供了手段。目前，对玉米、水稻和小麦等大宗农作物进行单一识别或两类间分类识别的技术研究较多，对研究区多种农作物同步分类识别的研究较少。本研究基于随机森林分类器利用Landsat 8数据开展宁夏农作物分类，对八种主要农作物春小麦、玉米、水稻、苜蓿、蔬菜、葡萄、枸杞和瓜类进行同步分类试验。结果表明：随机森林方法可以满足研究区内多类作物同步监测的需求，精度可达80%以上。单时相分类精度可达到81.8%，后分类处理精度可达到82.8%，时间序列分类精度可达到85.1%，时间序列分类和后分类处理可以有效提高分类精度。随机森林分类精度随着树数量的增加而增大，当树的数量足够多时，模型趋于稳定，特征变量对精度的影响被控制在一定范围内，当特征变量设置为总特征变量的平方根或对数时，精度达到最佳。因此，基于对分类实验时效性的考虑，将参数分别设置為Ntree= 100，Mtry=总特征变量的平方根或对数。

关键词：随机森林;多作物识别;分类后处理;单时相;时间序列

Multi - Crop Synchronization Identification Based on Random ForestXu Qjl，2 ， Li Qjliang1，2 ， Mathilde De Vroey3 ， Zhang Wupingl ， Fan Jinlong2

Abstract Identifying crop type with remotely sensed image is the fundamental step for calculating croparea and monitoring crop growth as well as estimating crop yield in the context of agricultural remote sensing.At present， there were many studies on single recognition or inter - classification recognition among the majorstaple crops， such as corn， rice and wheat， however， less researches have been done on simultaneous classifi-cation and recognition of multiple crops in the study area. Based on the random forest classifier and Landsat 8data， this study carried out the crop classification in Ningxia， the crops including rice， corn， wheat， clover，grapes， alfalfa， vegetables and melons. The results showed that it was feasible to use the random forest methodto simultaneously identify multiple crops in Ningxia， and the accuracy could reach more than 80%. Thesingle - phase classification accuracy could reach 81.8% ， after classification post - processing accuracy couldbe up t0 82.8% . The time series classification accuracy could reach 85.1% . Time series classification andpost - classification processing could effectively improve classification accuracy. The accuracy of random forestclassification increased with the increase of the number of trees. The model tended to be stable when the num-ber of trees was sufficient， and the influence of characteristic variables on the accuracv was controlled within acertain range. The precision was optimal when the feature variable was set to the square root or logarithm ofthe total feature variable. Therefore， based on the consideration of the timeliness of the classification experi-ment， the parameters were respectively set to Ntree = 100， and Mtry = the square root or logarithm of the totalcharacteristic variable.

Keywords Random forest; Multi - crop classification; Post - classification; Single - phase; Time series

掌握农作物种植的空间分布状况，对于宏观指导农业生产、合理采取作物管理措施具有重要的现实意义[1-3]。传统的农业统计报表和抽样调查，通常需耗费大量人力物力且持续时间长[4]，对当季农业生产的指导作用有限。遥感技术具有观测面积大、观测周期短、数据的综合性和可比性较强等特点，为及时、准确、客观地获得作物种植面积提供了技术手段和重要基础数据[5，6]。……