无题项A mapping framework for discrete biological soil crusts in degraded ecosystems using UAV multispectral sensing and ensemble learning目

生物土壤结皮 （BSC） 是生态恢复的先驱群落，对于退化的生态系统至关重要。然而，BSC 通常以具有复杂光谱混合的孤立斑块的形式出现，使得准确的映射变得困难。本研究引入了一种结合无人机多光谱遥感和集合学习的创新框架，以精确绘制恢复后磷石膏储藏地点离散的BSCs。获得了来自86个地点的综合无人机图像、PlanetScope多光谱卫星数据和现场采集的叶绿素a（Chla）样品。随机森林（RF）分类和集合回归相结合的结果表明，基于无人机的分类（精度0.62）明显优于基于卫星的方法（精度0.42）。继无人机之后，投票集合模型生成的多光谱Chla反演图达到了最高的预测精度（R² = 0.779）。使用Chla阈值进行重新分类，进一步将总体准确率提高到0.814，准确区分了土壤和石膏、BSC、凋落物、草本植物和木本植物5种土地覆盖类型。估计的 BSC 覆盖率与传统象限调查的覆盖率非常吻合。通过BSCs在磷矿尾矿池和稻田中的分布监测仪进一步验证了该方法的有效性，平均召回率为0.78。研究表明，结合高分辨率无人机遥感、Chla生化指标和集合学习，有效解决了退化生态系统中BSCs监测的挑战，为生态系统恢复评估提供了宝贵的支持。

Biological Soil Crusts (BSCs) are pioneer communities for ecological restoration and critical to degraded ecosystems. However, BSCs usually appear as isolated patches with complex spectral mixing, which renders accurate mapping difficult. This study proposes an innovative framework combining unmanned aerial vehicle (UAV) multispectral remote sensing and ensemble learning to accurately map discrete BSCs at restored phosphogypsum storage sites. Comprehensive UAV imagery, PlanetScope multispectral satellite data, and field-collected chlorophyll a (Chla) samples from 86 sites were obtained. The results combining Random Forest (RF) classification and ensemble regression showed that UAV-based classification (accuracy = 0.62) significantly outperformed satellite-based methods (accuracy = 0.42). Subsequent to UAV-based analysis, the multispectral Chla inversion maps generated by the voting ensemble model achieved the highest prediction accuracy (R² = 0.779). Reclassification using the Chla threshold further raised the overall accuracy to 0.814, accurately distinguishing five land cover types: soil and gypsum, BSCs, litter, herbaceous plants, and woody plants. The estimated BSC coverage agreed well with that from traditional quadrat surveys. The effectiveness of this method was further validated through monitoring BSC distribution in phosphorus ore tailing ponds and paddy fields, with an average recall rate of 0.78. This study demonstrates that combining high-resolution UAV remote sensing, Chla biochemical indicators and ensemble learning effectively addresses the challenges of BSC monitoring in degraded ecosystems, providing valuable support for ecosystem restoration assessment.