Unveiling fine-scale distribution of endemic shrub Prunus ledebouriana through integrating multi-source remote sensing with deep learning.

This study aims to unveil a fine-scale spatial distribution of endemic shrub Prunus ledebouriana (Schltdl.) Y.Y.Yao in Kazakhstan's Tarbagatay National Park by integrating multi-source remote sensing with deep learning. Accurate characterization of plant species distribution requires spatially precise ground-truth data; however, conventional GPS-based methods often introduce positional uncertainties that compromise alignment with very high-resolution imagery. To overcome this limitation, we employed a hybrid ground-truthing strategy that combines very high-resolution (5 × 5 cm) drone imagery with field-based onscreen digitization, enabling the generation of spatially accurate reference data. This data was used to extract training and validation data points from 18 predictor variables, encompassing spectral bands, vegetation indices, and texture features derived from Pléiades Neo imagery (30 × 30 cm), along with ancillary topographic and climatic variables. Based on these inputs, a deep one-dimensional convolutional neural network (1D CNN) model was developed to characterize the spatial distribution of P. ledebouriana. The model achieved an overall mapping accuracy of about 80%, with feature importance analysis highlighting texture metrics as the most influential predictors. Results revealed that P. ledebouriana covers about 7.5% of the study area; with distribution strongly linked to specific topographic settings. Nearly 70% of occurrences were found between 700 and 1200 m elevation, peaking at 900-1000 m.a.s.l., and about 75% were located on moderately slopped terrains (5-30%). Aspect also influenced distribution, with 83% of occurrences on southeast- to west-facing hillslopes. The limited occurrence of P. ledebouriana in lowland agricultural areas and on steep slope terrains suggests a combined influence of anthropogenic land-use pressures and ecological preferences. This study demonstrates the potential of integrating spatially precise ground-truthing, multi-source remote sensing, and deep learning for accurately mapping plant species distribution in mountainous drylands, supporting biodiversity monitoring and conservation planning in these fragile ecosystems.