基于决策树和混合像元分解的玉米种植面积提取方法

Extraction of Maize Planting Area Based on Decision Tree and Mixed-pixel Unmixing Methods

Landsat 8影像具有较高空间分辨率和时间分辨率,长时间序列Landsat 8-NDVI曲线反映农作物的物候历、种植模式和种植结构信息,是精确提取玉米种植面积的理想数据源。基于时序Landsat 8-NDVI影像提取玉米种植面积的方法中,决策树方法快速、高效,可通过多阈值限定进行分类,但由于混合像元问题,如果阈值设置过宽,提取面积偏大;阈值设置过窄,提取面积偏小;混合像元分解通过计算端元组分丰度可以排除异质地类干扰。因此,以时序NDVI为数据源、耦合使用2种算法是精确提取作物种植面积的有效方法。本研究基于时序Landsat 8-NDVI,提取河北省保定市大田玉米的种植面积。首先,分析典型作物区的NDVI曲线特征,并构建决策树从而初步提取早播夏玉米、小麦夏玉米和春玉米的分布范围。然后,根据端元平均NDVI波谱曲线,进行3种玉米混合度分解,进而根据玉米丰度比例精确提取玉米种植面积。精度评价结果表明:利用本方法提取的玉米种植区总分类精度在98%以上,Kappa系数在0.97以上;所提取的玉米种植类型主要是夏玉米,春玉米种植主要集中在涿州市中部,这与实地调查结果一致。上述定量和定性的评价结果表明该方法可用于快速、精确提取玉米种植面积。

Landsat 8 remote sensingimages possess higher spatial resolution and higher temporal resolution. The time-series Landsat 8 - NDVI metrics could reflect the phenology calendar, planting pattern, planting structure and planting area information due to its high spatial resolution and high temporal resolution, thus it is an ideal data source for accurate extraction of maize planting area. In most extraction methods, the decision tree classification method is considered to be rapid and efficient, which could extract maize planting area using multi-threshold. However, because of the mixed-pixel, both the larger and smaller threshold will lead to errors. This problem could be resolved by mixed-pixel unmixing method using endmember abundance calculation to eliminate the disturbance of heterogeneous classes. Therefore, taking time-series Landsat 8 -NDVI metrics as data source and using the combined method of decision tree and mixed-pixel unmixing methods are effective way to extract crop planting area. The maize planting area in Hebei Province was extracted in this paper based ontime-series Landsat 8 - NDVI. Firstly, the features of time-series Landsat 8 - NDVI curves were analyzed and the decision tree was built to get the distribution of early sowing maize, interplanted summer maize and spring maize. Secondly,mixedness decomposition was calculated among three kinds of maize based on mean NDVI spectral curve of endmember, so maize planting area could be extracted accurately by using computed maize endmember abundance. The accuracy assessment resuhs indicated that the overall classification accuracy of maize planting area was higher than 98% and Kappa coefficient was higher than 0.97. Generally speaking, the main planting crop was summer maize, and spring maize was mostly planted in the south part of Zhuozhou City. These results were accordant with field work data. The above quantitative and qualitative accuracy assessment results indicated that this method can be used to extract maize planted area quickly and accurately.