基于多时相遥感几何与光谱特征的水稻种植区识别

    Rice planting area identification based on multi-temporal remote sensing geometric and spectral features

    • 摘要:
      目的 针对复杂农田景观中耕地破碎、水塘混杂导致水稻遥感识别困难的问题,构建一套适用于该类区域的水稻识别方法,为市域耕地资源调查和作物种植结构监测提供技术支撑。
      方法 以鄂州市为例,首先基于高空间分辨率遥感影像,结合田块几何纹理结构使用深度学习方法开展耕地地块级分割;并引入归一化差值植被指数(Normalized difference vegetation index,NDVI)和归一化差值水体指数(Normalized difference water index,NDWI)光谱特征剔除水塘干扰。随后,基于水稻的生长变化规律,在耕地范围内构建多时相的光谱特征集,分别采用随机森林(Random forest,RF)、极端梯度提升(Extreme gradient boosting,XGBoost)和长短期记忆网络(Long short-term memory,LSTM)模型进行水稻识别。
      结果 剔除水塘干扰后,耕地提取的精确率达到95.1%。在水稻种植区识别中, XGBoost模型表现最优,总体精度达到92.86%,同时SHAP分析表明水稻生育期内植被指数的时序变化特征是识别的重要依据。
      结论 基于多时相遥感影像的几何光谱特征能够有效提升复杂农田景观下的水稻识别效果。

       

      Abstract:
      Objective To address the challenge in identification of rice planting areas via remote sensing in complex farmland with fragmented cropland and interspersed ponds, this study develops a method for identifying rice planting areas in such regions, supporting municipal cultivated land surveys and crop planting structure monitoring.
      Method Taking Ezhou City as an example, this study first employs deep learning for parcel-level cropland segmentation using high-spatial-resolution remote sensing imagery based on geometric texture of field parcels, and introduces normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) to eliminate interference from ponds. Then, a multi-temporal spectral feature set is constructed within cropland areas based on rice growth patterns, and identification of rice planting areas is performed using RF, XGBoost, and LSTM models.
      Result After removing pond areas, cropland extraction precision reached 95.1%. Among rice planting area identification models, XGBoost performed best, with an overall accuracy of 92.86%. SHAP analysis further indicated that temporal variations in vegetation indices during the rice growing season are an important basis for identifying rice planting areas.
      Conclusion The geometric and spectral features from multi-temporal remote sensing imagery can effectively improve identification of rice planting areas in complex farmland landscapes.

       

    /

    返回文章
    返回