基于RF算法优选多时相特征的冬小麦空间分布自动解译

Automatic Interpretation of Spatial Distribution of Winter Wheat Based on Random Forest Algorithm to Optimize Multi-temporal Features

摘要:

为探讨如何利用遥感影像自动解译技术，实现冬小麦种植情况统计调查、提高提取精度，选择冬小麦关键生育期6个时相的高分二号遥感影像数据，分别从6个时相的近红外灰度（NIR）、红波段灰度（R）、绿波段灰度（G）、蓝波段灰度（B）、比值植被指数（RVI）、归一化植被指数（NDVI）6个特征中优选出对冬小麦面积提取最敏感的1个特征作为输入变量，每个时相选择1个特征，6个时相共选出6个特征作为输入变量，利用随机森林算法构建模型，提取冬小麦空间分布特征。选择研究区不同长势、不同种植品种的地块样本构建训练集，利用多时相特征构建模型，并将模型推广应用于整个大厂回族自治县，得到大厂回族自治县冬小麦的空间分布情况。通过与统计结果对比分析，经过多时相特征优选构建的模型对冬小麦的识别精度接近90%。经过样本优化和后期处理仍可提升精度，此方法能在保证提取精度的前提下对冬小麦进行快速提取，提高相应的工作效率。

Abstract:

In order to explore how to use the remote sensing image automatic interpretation technology to realize the winter wheat planting statistics survey and improve its extraction accuracy，the Gaofen-2 remote sensing image data of six key growth periods of winter wheat were selected. One of the most sensitive features to winter wheat area was selected respectively as the input variable from six features of near-infrared gray (NIR), red band gray (R), green band gray (G), blue wave band gray (B), ratio vegetation index (RVI) and normalized difference vegetation index (NDVI). One feature was selected for each time phase, and six features were selected for the six time phases. A model was constructed by using the random forest algorithm to extract winter wheat. The training set was constructed by selecting land samples with different growth and planting varieties in the study area. The model was constructed based on the multi-temporal features and applied to the whole Dachang Hui Autonomous County. The spatial distribution of winter wheat in Dachang Hui Autonomous County was obtained. Compared with the statistical results, the recognition accuracy of the model constructed by multi-temporal feature optimization was close to 90%. After sample optimization and post-processing, the accuracy can still be improved. This method can quickly extract winter wheat on the premise of ensuring the extraction accuracy, and greatly improve the corresponding work efficiency.