基于Landsat8 OLI与MODIS数据的洪涝季节作物种植结构提取

洪涝灾害会造成农作物严重受损,因此洪涝季节作物的种植结构是估算洪涝灾害损失、进行防灾减灾措施的必要信息。为了能够快速便捷地提取洪涝季节作物种植结构,该文以湖北省监利县为研究区域,探讨了采用空间分辨率较高的Landsat8陆地成像仪(operational land imager,OLI)影像和时间分辨率较高的中分辨率成像光谱仪(moderate-resolution imaging spectroradiometer,MODIS)数据,综合利用多源多时相遥感影像提取中小尺度范围的洪涝季节作物种植结构的方法。首先利用MODIS数据建立作物的归一化植被指数(normalized difference vegetation index,NDVI)时间序列曲线,并采用改进后的Savitzky-Golay滤波器对曲线进行平滑处理,然后根据作物的物候特征设定阈值,界定作物种类,进而以此为依据在作物关键生育时期的Landsat8 OLI高清影像中选择合适的感兴趣区域(region of interest,ROI)作为先验知识,使用BP(back propagation)神经网络模型对OLI数据进行监督分类,提取作物种植面积分布。最后利用统计数据与资源三号卫星数据对提取结果进行验证,平均精度达到88%,能够较准确地反映监利县洪涝季节作物的分布情况。该研究可为洪涝灾害损失估算提供可靠基础。

Extraction of crop planting structure in seasons prone to waterlogging using Landsat8 OLI and MODIS data

Abstract: Flood disaster occurs frequently in China and brings severe disaster to crops. Therefore, the crop plant structure in seasons prone to waterlogging becomes significant information for studies on flood loss, flood control, and disaster mitigation. Under these conditions, this paper presents a fast and convenient method to extract the crop plant structure in small scale areas during seasons prone to waterlogging, based on multi-sensor and multi-temporal remote sensing data. Landsat8 OLI and MODIS data were chosen because of the advantages such as it being free of charge and easy to search for and download. These two types of data showed the characteristics of crops' growth respectively in space and time, leading to a proper combination for crop planting structure extraction. If one only uses MODIS data to build the extraction model, the spatial resolution is too low to get the planting structure in small scale areas. On the other hand, just classifying the OLI images by visual interpretation sometimes could not determine the types of crops. The Jianli County in Jingzhou City, Hubei Province was chosen as the study area. The seasons prone to waterlogging in Jianli mainly include June, July, and August, and are related to crops such as early-season rice, middle-season rice, late rice, and cotton. Here are the extracting models for the four major kinds of crops: NDVI value of cotton grows to the peak in early July and stays high until September; early-season rice NDVI maximum value appears in middle June, and it becomes completely mature in late July; middle-rice NDVI value starts to increase at the end of May and reaches its peak in early or middle July before falling to a decline; NDVI value of late rice goes up in mid-July and the peak value appears in middle or late August, and then the value begins to decline. First, a time series curve of NDVI was built from the MODIS data, which was later smoothed by an improved Savitzky-Golay filter. The improved Savitzky-Golay filter reserved the authenticity of data at both ends of the NDVI time series while further improving the smoothness of the curve. To distinguish the types of crops, threshold values of NDVI for different crops were set according to corresponding phenophases. Based on the characteristics and threshold values of NDVI time series, appropriate ROIs (Region of Interest) in the Landsat8 OLI images in key growth stages of different crops were selected as prior knowledge for training. Finally, the area and distribution of the four studied crops were extracted by a BP Neural Net Supervised Classification. The experimental results agreed well with the statistical data and a ZY-3 image which had a spatial definition of 6 meters, and obtained an average precision above 90%. It was concluded that the proposed method in this paper is simple and easy operating. Moreover, it accurately reflected the real situation of crop distribution in the Jianli area, and is suited for extraction of the plant structure in small scale areas like Jianli. Therefore, this method provides a reliable basis for related research studies on flood disaster.