Drone remote sensing and machine learning for green stormwater infrastructure condition assessment

Maintenance and operations of green stormwater infrastructure is critical to preserve the functionality of urban stormwater infrastructure. However, doing so is a challenge due to the disperse locations of green stormwater infrastructure that may be difficult to access, which results in limited and inconsistent inspections that are also human and resource intensive. The objective of this study is to overcome this limitation through a novel approach to green stormwater infrastructure inspection that applies machine learning models to remote sensing data from an unmanned aerial system to assess green stormwater infrastructure landcover. To do so, machine learning models were applied to categorize land cover of green stormwater infrastructure into 4 condition-related classes: healthy plants, unhealthy plants, dead plants and organic material, and inorganic material. Models were trained and tested via multitemporal analysis at 12 unique locations encompassing various green stormwater infrastructure types (e.g., bioswale, green roof, rain garden, native planting area). The landcover classification accuracy assessments showed that supervised object-based and pixel-based methods exhibited similar overall accuracy (87 % and 88 %, respectively) during training and testing. Notably, Random Trees and Support Vector Machine algorithms outperformed Maximum Likelihood and k-Nearest Neighbors by an average of (+4 %). Overall, these methods can be used to obtain informative data that can enhance green stormwater infrastructure monitoring and maintenance efforts.