Assessment of herbaceous vegetation classification using orthophotos produced from the image acquired with unmanned aerial systems

Arguably the most popular remote-sensing products are classified images. However, there are no definitive procedures to assess classification accuracy that simultaneously consider resources available and field efforts. The explosive usage of unmanned aerial systems (UAS) in land surveys adds new challenges to classification assessment, as orthorectified images usually contain significant artifacts. This study aims to identify the optimal ratio between training and validation sample size within a supervised classification approach applied to UAS orthophotos. As a case study, we used a wetland area west of Portland, OR, USA, treated with various glyphosate formulations to control Phalaris arundinacea, commonly known as reed canary grass. A completely randomized design with five replications and six glyphosate formulations was used to assess P. arundinacea vigor following repeated herbicide applications. The change in P. arundinacea vitality was monitored with high-resolution four-band imagery acquired with a SlantRange 3PX camera installed on a DJI Matrice 210. The orthophotos created from images were produced with Pix4D, which was subsequently preprocessed with ERDAS Imagine 2020 to reduce the noise, shadows, and artifacts. All images were classified with the maximum likelihood classification algorithm. Simple random and stratified random sampling methods were applied to collect training and validation samples, evaluating eight ratios of training to validation samples to assess their classification accuracy. We found that increasing the training-to-validation sample size ratio enhances accuracy, with the 3:1 ratio being the most reliable in classifying P. arundinacea vigor. Our study provides evidence that image preprocessing and enhancement are essential for UAS-based imagery.