Comparing seasonally-varied CNNs in vegetation segmentation performance
MetadataShow full item record
In the Netherlands, management of coastal dunes has changed focus from short-term flood prevention and drinking water protection to long-term biodiversity promotion and back-dune development. To understand the impact of dune management activities, efficient monitoring methods are required. The use of unmanned aerial vehicles (UAVs) to provide images of areas of interest, and convolutional neural networks (CNNs) to segment these images, are promising developments in remote sensing. Additionally, human-in-the-loop machine learning (HITLML) provides an efficient way to annotate remotely sensed images for land cover classification. Vegetation phenology has shown to impact identifiability from aerial images in some applications. However, the viability of the combined application of these four elements (UAVs, CNNs, HITLML and phenology) to coastal dune monitoring is yet to be assessed. Here we show that these elements can potentially be effective in monitoring coastal dune development. We found that phenology impacts CNN performance, and CNNs trained on multi- season data perform more consistently than single-season CNNs. Additionally, increasing training data volume does not always improve CNN performance. Overall multi-season models’ performance (measured by f1 score) was around 5% better than season-specific CNNs, which is in line with findings of other research. Predictions on specific seasons revealed roughly equivalent performance by season- specific and multi-season CNNs. The impact of increasing the dataset size had minimal effect on model performance. White dunes were most error-prone of all habitat types, being most frequently incorrectly predicted as grey dunes. Our results demonstrate that in the context of dune habitat analysis, training data should comprise seasonal diversity, and that beyond a certain point, increasing the sample size (without additional temporal diversity) is likely to have a limited effect. We anticipate our research to lead to improved practices in automated dune monitoring. Diversity of data appears to have a greater impact on model performance than volume alone, and CNNs provide a useful way to create habitat maps for automated dune monitoring. Additional research should incorporate other factors, such as the impact of weather conditions on results, as well as using ensemble models.