Popis: |
Gullies and cliff erosion are significant indicators of land degradation. Knowledge of the spatial distribution and dimensions of these erosion features is needed to effectively quantify sediment budgets and to implement erosion mitigation measures. Expert delineation can help identify features at a local sale, however, mapping larger extents becomes time consuming. Object detection techniques based on aerial photographs and LiDAR elevation data can improve the automated delineation of such features. In this study, we tested a region-based convolutional neural network (Mask-RCNN) deep learning approach to identify gully and cliff features. An expert-based delineation of gully and cliff features was performed in the Wairoa catchment in Hawke’s Bay, New Zealand based on aerial photographs obtained between 2017 and 2020. These delineations served as reference data to create labelled chips for training deep learning models. Several terrain derivatives from the LiDAR digital elevation model (DEM), including slope, hillshade, slope length and steepness (LS) factor, and terrain ruggedness index, were computed. The terrain derivatives and spectral bands (R-G-B-NIR) from the aerial photographs were used to train deep learning models based on different band combinations. Despite achieving high accuracy (average precision score above 85%) during training, transferring the models to validation areas resulted in low detection rates, with a large number of false negatives. However, the correctly detected erosion features correspond very well to the reference data delineations, even achieving good results for the differentiation between gullies and cliffs. A closer inspection of the false positive features suggests that the reference data could be incomplete. Our study shows that deep learning has a high potential for automated gully and cliff mapping, but further improvement of the model transferability is needed. A combination of automated and expert-based delineation would potentially result in reliable and efficient erosion feature detection. |