【摘要】 针对水产养殖遥感分类问题，该文提出了一种基于光谱时序特征的精细分类方法，以福建省三沙湾为研究区，综合利用吉林一号影像和2023—2024年Sentinel-2时间序列数据，构建水产养殖样本集、NDAI、NDVI、NDAWI和NDCI四种光谱指数特征及指数融合特征的时序数据集，利用RF、SVM、GBDT和XGBoost四种机器学习算法和五种特征分别对海带、龙须菜和网箱3种水产养殖进行分类。结果表明：(1)单指数时序特征中NDAI表现最优，多指数融合时序特征显著提升分类精度，其中XGBoost基于融合特征的分类效果最佳，F1分数为96.04%;(2)典型区域对比显示，融合特征可有效增强养殖边界识别清晰度、减少误识漏提；(3)三沙湾水产养殖呈现“海带条带状集中分布、龙须菜广域覆盖、网箱分散嵌于藻类养殖区”的空间分异格局。更多还原

【Abstract】 This paper proposes a fine classification method based on spectral temporal features for remote sensing classification of aquaculture. Taking Sansha Bay in Fujian Province as the research area, the Jilin-1 image and Sentinel-2 time series data from 2023 to 2024 are comprehensively utilized to construct a time-series dataset of aquaculture samples, four spectral index features including NDAI,NDVI,NDAWI,and NDCI,and index fusion features. RF,SVM,GBDT,and XGBoost machine learning algorithms and five features are used to classify three types of aquaculture, namely kelp, dragon’s beard vegetable, and net cage. The results show that:(1)among single-index time-series features, NDAI performed the best, while the fused spectral time-series features significantly improved classification accuracy, with XGBoost achieving the highest F1-score of 96.04%;(2)comparative analysis in typical regions indicates that the fused features enhance boundary clarity, reduce misclassification and omission;(3)the spatial distribution pattern of aquaculture in Sansha Bay is characterized by“striped kelp farming in concentrated zones, wide-area gracilaria coverage, and scattered cages embedded within seaweed cultivation areas.”更多还原