基于无人机高光谱影像的建筑垃圾分类研究

建筑垃圾“围城”已经成为现阶段城市环境治理面临的主要问题，严重制约了城市生态环境的可持续发展，做好建筑垃圾的分类对保护城市水资源、提高城市土地利用率、提升居民生活质量意义重大。该研究将GaiaSky-mini 2推扫式机载高光谱成像仪（400～1 000 nm）搭载在经纬M600Pro无人机上，选择晴朗无风的试验环境，实时获取研究区高光谱遥感影像。对采集的研究区高光谱遥感影像进行几何校正、图像裁剪、辐射校正等预处理；将研究区内地物分为背景地物和建筑垃圾两大类，其中背景地物包括芦苇、蒿子、水体、阴影、裸土和柏油路，建筑垃圾包括白色塑料、防尘布、地基渣土和瓦砾砂石；基于影像像元选取样本点，分别提取研究区内6种背景地物和4种建筑垃圾的光谱信息，制作光谱曲线，并依据光谱特征差异，选取特征波段，通过波段计算统计并选取合理阈值，利用决策树分类法实现背景地物的分离和建筑垃圾的识别提取；针对不同类别的背景地物和建筑垃圾分别选取验证样本点，对背景地物的分离结果和建筑垃圾的识别结果进行精度评价。结果表明，背景地物和建筑垃圾总体识别精度为85.91%，Kappa系数为0.845；针对建立的背景地物分离决策树，6种背景地物的分类效果均较好，其中芦苇、柏油路和裸土的生产者精度为95%，整体能较好的将背景地物分离；针对建立的建筑垃圾识别决策树，防尘布和瓦砾砂石的生产者精度为95%，白色塑料和地基渣土的生产者精度为90%，能精确的提取研究区内的建筑垃圾。研究表明决策树分类法在无人机高光谱遥感影像中实现建筑垃圾的识别与提取具有很好的分类准确度，同时也验证了无人机高光谱遥感在建筑垃圾分类提取领域的科学性和可行性，对未来建筑垃圾的分类识别工作具有一定的实际意义。

Research on Classification of Construction Waste Based on UAV Hyperspectral Image

The “siege” of construction waste has become the main problem of urban environmental pollution at this stage, severely restricting the sustainable development of the urban ecological environment. A good classification of construction waste is of great significance to protecting urban water resources, improving the utilization rate of urban land and improving residents’ quality of life. In this paper, the GaiaSky-mini 2 push-broom airborne specular imager (400~1 000 nm) is mounted on the DJ MATRICE M600Pro UAV, and a clean and windless test environment is selected to collect hyperspectral remote sensing images of the study area in real-time. The hyperspectral remote sensing images of the study area were preprocessed by geometric correction, image cropping and radiometric correction; The objects in the study were divided into two categories: background objects, including reed, wormwood, water, shadow, bare soil and asphalt road, and construction waste including white plastic, dust cloth, foundation residue and rubble sand. Based on pixel points, select the regions of interest (ROI) of various features as training samples, extract the spectral information of six background features and four types of construction waste in the study area, and make spectral curves based on different spectral feature differences between features. Select feature bands, calculate statistics through bands and select reasonable thresholds, use decision tree classification to separate background features and identify and extract construction waste in the study area. Target different background features and construction waste types were selected to verify the sample points and evaluate the accuracy of the separation results of background features and the identification results of construction waste.The results show that the overall recognition and classification accuracy of background features and construction waste is 85.91%, and the Kappa coefficient is 0.845. According to the established decision tree for the separation of background features, the classification effect of six background features is good, among which the producer accuracy of reed, asphalt road and bare soil is 95%, and the overall separation of background features is good. According to the established construction waste identification decision tree, the producer accuracy of dust cloth and rubble sand is 95%, and the producer accuracy of white plastic and foundation residue is 90%, which can accurately extract construction waste in the study area. This study shows that decision tree classification is realized in the unmanned aerial vehicle (UAV) hyperspectral remote sensing image recognition and extraction of the construction waste has good classification accuracy. Moreover, to verify the unscrewed aerial vehicle (UAV) hyperspectral remote sensing in the field of construction waste classification to extract the scientific nature and feasibility of construction waste classification recognition for future work has a specific practical significance.