Woody vegetation on agricultural land provides various ecosystem services, such as conserving biodiversity, preventing erosion, storing carbon and helping to control pests. However, data on its extent are limited. Therefore, we tested the feasibility of detecting it using remote sensing data, specifically the national aerolaser scanning data, which was used to create a canopy height model, and the national orthophoto. Supervised classification methods were used to detect the following woody vegetation classes: single trees, trees in lines, trees in orchards and olive groves, groups of trees and bushes, hedges, riparian vegetation and forest. Initially, classification was performed using convolutional neural networks (DeepLabV3+, HRNet and U-Net), which successfully classified woody vegetation overall but were less successful at differentiating between the various woody vegetation classes. Therefore, we also tested the detection of woody vegetation using object classification, which provided the highest mean Jaccard index in the test areas (44.52%). This was achieved through a process involving binary classification of tree canopies using HRNet and multiclass classification of woody vegetation classes using a multilayer perceptron. This process was used to map woody vegetation throughout Slovenia. The highest density of woody vegetation features is in the Littoral region, and the lowest density is in the Alpine region. Groups of trees and bushes are the most common class in most landscapes, with the exceptions being landscapes near the coast, where orchards and olive groves are numerous, and karst landscapes, which have relatively many hedges. Additionally, landscapes adjacent to larger rivers are notable for their abundant riparian vegetation. The results are less accurate in the high mountains than elsewhere due to a lack of training data from such areas.
|
