基于无人机多光谱和SOLO模型的防护林枯树提取方法

Extraction of Dead Trees from Shelter-Forest Based on UAV Multispectral and SOLO Model

针对三北防护林树种混交比例低、灌水不足、空间分布不均匀、病虫害时发造成的防护林衰退问题,提出了一种小样本下基于无人机多光谱波段生成的SOLO(segmenting objects by locations)实例分割提取枯树的方法.通过无人机搭载Micro MCA12 Snap多光谱相机获取高空间分辨率影像,将多光谱可见光波段(波段5、波段3、波段1)导出并标注样本,基于最佳指数因子选出排名前10的波段组合进行数据集的扩充,最后基于实例分割SOLO模型进行防护林枯树的提取.实验结果表明,加入健康树样本,基于ResNet-101+FPN的SOLO模型AP从61.3%提升到63.8%,ResNet-50+FPN组合AP从60.7%提升到63.6%,同时进一步验证了这种样本增强方式的有效性.

Aiming at the decline of three-north shelter-forest caused by low mixed proportion of tree species, insufficient irrigation, uneven spatial distribution and frequent occurrence of diseases and pests, a method of extracting dead trees based on SOLO generated by UAV multi-spectral band under few-shot is proposed in this paper. The UAV is equipped with a Micro MCA12 Snap multispectral camera to obtain high spatial resolution images, derive and labele samples of multispectral visible bands(band 5,band 3,band 1). Based on the selected top 10 band combinations according to the best exponential factor, the dataset is expanded, and finally dead trees of shelter-forest are extracted based on the instance segmentation SOLO model. The experimental results show that with the addition of health tree samples, the SOLO model AP based on ResNet-101+FPN is increased from 61.3% to 63.8%,and the combined ResNet-50+FPN AP is increased from 60.7% to 63.6%. At the same time, the effectiveness of this sample enhancement method is further verified.