Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

< Back to Article

Fig 1.

Location of the study area.

A: Landscape of outstanding features Subotička peščara. B: Satellite image of the sandy steppe habitat Sunčani salaš, C: overlapped with an UAV orthophoto mosaic from 09/30/2021.

More »

Fig 1 Expand

Fig 2.

An overview of research workflow.

All activities were grouped around three main topics shown in the upper part of the diagram. Next to each functional block are the names of the corresponding Tables and Figures in the text. E.g. Tables 8 and 9 summarize the main results of the proposed vegetation study and the performances of the designed UAV time series classification models.

More »

Fig 2 Expand

Table 1.

Data collection dynamics.

More »

Table 1 Expand

Fig 3.

Orthmosaics of the test site illustrating seasonality effect.

A: April until September. B: Time lapse of the same spot from which UAV campaigns were initiated.

More »

Fig 3 Expand

Fig 4.

Locations of 1m x 1m vegetation sampling plots and the transect line.

According to Table 9 assigned habitat types are: 1–Young steppe I (E1.2C1); 2–Young steppe II (E1.2C2); 3–Bare / fallow land (I1.51); 4–Young steppe I (E1.2C2), while the Forest Steppe I (G1.4) was sampled along the transect line (violet). Small blue dots represent shrubs. Five distinct observation Zones (I–V) are marked in brown.

More »

Fig 4 Expand

Fig 5.

Spatial details in the acquired UAV images.

Same part of the scene at different zoom levels (image from the end of June).

More »

Fig 5 Expand

Fig 6.

Sampling of the ground-truth image data.

A: For class C0 or “Steppe”, each pixel inside the inscribed circle was used as a source of labeled data. B: In the case of C1, C2, and C3, only the point sampling of the central GPS position was performed.

More »

Fig 6 Expand

Table 2.

Proposed GLCM texture descriptors.

Computed for 3 different scales (analysis window sizes) by averaging over the pixel pairs at distance of and positioned at 45° relative to each other. In total 54 FPP.

More »

Table 2 Expand

Fig 7.

Proposed texture and morphological image descriptors.

A: Comparison of three GLCM texture features: entropy, correlation, and homegeneity. Scatter plots indicate good discriminability between spatial neihbourhoods of the pixels corresponding to different types of vegetation patterns (patches denoted as sets A and B in the upper-right part of the figure). B: Morphological profile (MP) features, corresponding to MP derivatives of the morphological opening and closing by reconstruction, as compared to the GLCM entropy feature maps of the same image area (GLCM entropy variants corresponding to different filter sizes).

More »

Fig 7 Expand

Table 3.

Proposed morphological descriptors.

Morphological opening and closing by reconstruction at σ = 1 different scales (filter sizes) with ball shaped filter of radius ∼ pixels and step size of , . In total 11 FPP.

More »

Table 3 Expand

Table 4.

Proposed spectral descriptors.

Computed from 3-channel RGB images according to the provided references. In total 9 FPP.

More »

Table 4 Expand

Table 5.

Identified observation zones and their characteristics.

More »

Table 5 Expand

Table 6.

Site specific steppe habitat types and their quality in relation to the EGS.

More »

Table 6 Expand

Table 7.

Identified steppe development phases on the grassland surface of Sunčani salaš.

More »

Table 7 Expand

Table 8.

Proposed indicators for monitoring steppe development phases.

More »

Table 8 Expand

Table 9.

Overview of the main habitat types and vegetation characteristics at the site.

More »

Table 9 Expand

Table 10.

Results of classification experiments for different high resolution descriptors.

a –“Steppe”, –“Shrubs”, –“Forest-steppe”, –“Bare/fallow land”.

More »

Table 10 Expand

Fig 8.

Visual comparison of image classification results.

A: Map produced by the model combining all textural, morphological and spectral descriptors over collected UAV image time series (in total 444 combined features per pixel). B: Improved map produced by the model based on only 15 out of 444 features (list of features selected by the VSURF procedure is shown on the right-hand side, please see the Methods section).

More »

Fig 8 Expand

Fig 9.

Classifier decisions with overlaid confidence scores.

Pixels with low confidence scores ( < 0.6) are shown in white. A: Map produced by the model based on combined image descriptors, without feature selection, and B: Model trained on selected 15 features. C: Side-by-side comparisons for the image details shown by the two true colour images on the left.

More »

Fig 9 Expand