基于多指数递进耦合的设施农业大棚识别方法

Progressive Coupling of Multiple Greenhouse Indices for Greenhouse Identification in Facility Agriculture

摘要:

准确获取农业大棚的空间分布信息对于保障粮食安全、优化设施农业布局和推动乡村振兴具有重要意义。为了实现大范围、自动化的大棚识别与监测,基于Sentinel – 2影像数据,以山东省寿光市为研究区,提出了一种基于光谱纹理特征融合的样本自动生成与筛选方法。首先,融合APGI、RPGI、PMLI、EWI、NDVI及缨帽变换特征,结合Cohen’s d效应量加权阈值分割生成潜在大棚样本;进一步引入旋转不变局部二值模式(Local binary pattern,LBP)纹理特征进行筛选,获得高置信度自动样本。随后,从样本分布一致性与分类性能两方面评价其可靠性,并结合随机森林(Random forest,RF)、支持向量机(Support vector machine,SVM)和梯度提升决策树(Gradient boosting decision tree,GBDT)开展分类试验。结果表明,样本容量为150时自动生成样本与人工样本一致性最佳;GBDT分类效果最优,总体精度达到95.0% ,与人工样本训练结果接近。海曙区扩展试验中,总体精度和Kappa系数分别达到89.2%和0.83,说明该方法在复杂分散大棚场景下仍具有较好适用性。与Global-PCG-10对比发现,本文结果在寿光市与该数据集具有较高的空间一致性,交并比为0.728;在海曙区复杂场景中交并比为0.290。该方法在显著降低人工标注成本的同时仍能保持较高的分类精度,为设施农业自动化监测及样本自生成研究提供了可行路径和技术支撑。

Abstract:

Accurate acquisition of the spatial distribution of agricultural greenhouses is of great significance for ensuring food security, optimizing the layout of facility agriculture, and promoting rural revitalization. To achieve large-scale and automated greenhouse identification and monitoring, an automatic sample generation and screening method was proposed based on spectral-textural feature fusion using Sentinel – 2 imagery, with Shouguang City, Shandong Province, as the study area. Firstly, APGI, RPGI, PMLI, EWI, NDVI, and tasseled cap transformation features were integrated, and potential greenhouse samples were generated through weighted threshold segmentation based on Cohen’ s d effect size. Rotation-invariant local binary pattern ( LBP) texture features were then introduced for further screening to obtain high-confidence automatically generated samples. Subsequently, the reliability of the generated samples was evaluated from the perspectives of sample distribution consistency and classification performance, and classification experiments were conducted by using RF, SVM, and GBDT. The results showed that the automatically generated samples achieved the highest consistency with manually labeled samples when the sample size was 150. GBDT yielded the best classification performance, with an overall accuracy of 95.0% , which was close to that obtained by using manually labeled samples. In the extended experiment conducted in Haishu District, the overall accuracy and Kappa coefficient reached 89.2% and 0.83, respectively, indicating that the proposed method maintained good applicability in complex and spatially scattered greenhouse scenarios. Comparison with Global-PCG-10 showed that the proposed results exhibited high spatial consistency with the dataset in Shouguang, with an intersection over union (IoU) of 0.728, whereas the IoU decreased to 0.290 in the more complex scenario of Haishu District. The proposed method significantly reduced the cost of manual labeling while maintaining high classification accuracy, providing a feasible pathway and technical support for automated monitoring of facility agriculture and research on self-generated samples.