不同花期冬油菜指数分类的敏感性

Sensitivity of Winter Rapeseed Classification Using Winter Rape Index across Different Flowering Stages

关键词:  冬油菜指数  光谱特征  长江中游  Sentinel-2  数据缺失

Key words:  Winter Rapeseed Index  Spectral characteristics  Middle reaches of the Yangtze River  Sentinel-2  Data missing

摘要:

冬油菜与冬小麦作为我国南方主要的冬季作物,其物候期光谱特征高度相似,主流遥感分类方法依赖盛花期两种冬季作物显著的光谱特征进行识别。然而,受南方地区多云雨天气影响,盛花期影像往往缺失,导致依赖关键物候期的冬油菜识别方法面临严重的弱特征分类难题。为此,本文以长江中游地区为研究区域,利用适用于冬油菜提取的冬油菜指数(WRI, Winter Rapeseed Index),并构建基于贝叶斯网络的冬油菜弱特征遥感分类与敏感性评估框架,旨在定量评估不同花期WRI对两类冬季作物的分类敏感性。基于混淆矩阵与显著性检验的性能评价结果表明:在盛花期,WRI及传统植被指数均具有较高的可分离性；在初花期和末花期等弱特征时相,传统指数的可分离性大幅下降,而WRI仍能保持较高的可分离度。基于贝叶斯网络的分类研究表明,本文方法可有效挖掘弱特征,使初花期和末花期的冬油菜提取精度分别达到90.57%和88.61%。本研究有效拓宽了冬油菜遥感识别的时相窗口,可在关键物候期数据缺失的背景下,发掘潜在非关键物候期的替代价值,实现冬油菜的有效提取。这一方法为中国南方地区多云雨、遥感数据缺失背景下的冬油菜空间分布提取提供了新的解决方案,具有重要的现实意义。

Abstract:

Winter rapeseed and winter wheat are the dominant winter crops in Southern China, sharing highly synchronized phenological stages and similar spectral characteristics. Traditional remote sensing classification relies heavily on the distinct yellow spectral features of winter rapeseed during the peak flowering stage. However, due to geographical differentiation and frequent cloud cover and rainfall in the subtropical region, it is often challenging to acquire high-quality images precisely at the peak flowering stage, posing a significant "weak-feature classification" challenge for winter rapeseed recognition during non-peak stages. In this study, taking the middle reaches of the Yangtze River as the experimental region, we applied the Winter Rapeseed Index (WRI) and constructed a weak-feature remote sensing classification and sensitivity assessment framework based on Bayesian Networks. The objective was to quantitatively evaluate the inter-class separability and sensitivity of WRI between the two winter crops across different phenological windows, including the early, peak, and late flowering stages. Performance evaluations incorporating confusion matrices and statistical hypothesis testing ($Z$-test) indicated that during the peak flowering stage, both WRI and traditional vegetation indices exhibited high separability between winter rapeseed and winter wheat. Notably, during the weak-feature phases (early and late flowering stages), the separability of traditional indices degraded substantially, whereas WRI maintained a high level of separability (Separability Index > 0.7). Classification results based on the Bayesian Network demonstrated that the integration of WRI effectively mined potential weak features, achieving classification accuracies of 90.57% and 88.61% for the early and late flowering stages, respectively. This study successfully expands the temporal window for winter rapeseed mapping, unlocking the substitute value of non-peak phenological phases when critical peak-flowering data is missing. The proposed methodology provides a robust and efficient solution for accurate winter rapeseed mapping in Southern China, where phenological spatial variation and image acquisition uncertainty are significant constraints.