Objective As a vital component of the global ecosystem, forests play an irreplaceable role in maintaining the global carbon balance and regulating climate. In recent years, intensified climate change has led to a significant increase in the frequency and intensity of forest fires, severely threatening the structure and function of forest ecosystems. How to achieve ecological restoration of burned areas through scientific governance and intelligent technologies has become a focus in both academia and practice. Current research on forest burned areas has evolved from early field surveys to satellite remote sensing in terms of data acquisition. Field surveys yield accurate data but are costly and limited in scope; satellite remote sensing offers large-scale monitoring capabilities but lacks the resolution to support detailed forestry investigations at medium and small scales. Thus, a trade-off between “precision” and “efficiency” persists in recent years, the advent of UAV remote sensing has provided new opportunities for burned area research. However, existing interpretation methods largely focus on characterizing “burn severity,” lacking diagnosis of “restoration potential” and “ecological needs,” and fall short in revealing the succession dynamics of key species such as pioneer and constructive species. Meanwhile, traditional machine learning models perform poorly in vegetation semantic segmentation in complex environments. In summary, the core bottleneck in current burned area ecological restoration research lies in the disconnect between the theoretical demand for in-depth understanding of site heterogeneity, ecological processes, and succession dynamics, and the lack of high-precision, high-efficiency spatially explicit data from existing technical means. To address this, this study conducts intelligent interpretation research aiming to providing digital and intelligent technical support for the ecological restoration of forest burned areas.

Methods Taking the burned area of Jinyun Mountain in Chongqing as a case study, this research integrates UAV multispectral remote sensing, deep learning, and GIS spatial analysis technologies to interpret site ecological characteristics from two dimensions: land cover type identification and vegetation recovery status analysis. For land cover identification, a multispectral UAV was used to acquire centimeter-level imagery. The normalized difference vegetation index (NDVI) was calculated based on post-fire spectral feature changes to preliminarily distinguish bare land, burn residues, and healthy vegetation. Subsequently, the U-Net deep learning model, suitable for pixel-level segmentation of high-resolution imagery, was employed and trained using data augmentation strategies to achieve high-precision automated classification of three key vegetation recovery types: trees, ferns, and bamboo. For vegetation recovery condition analysis, slope, aspect, and slope position factors were extracted from a high-resolution DEM and classified according to the LY/T 3315-2022 Technical Regulations for Forest Site Quality Evaluation. Through coupled analysis, areas with similar light, temperature-moisture, and nutrient conditions were integrated into geographic restoration units. On this basis, the Strahler stream order method was used for flow accumulation (FA) analysis to reveal differences in hydrological conditions and soil erosion risks among units. Integrating the above information, the ecological needs and resource endowments of each unit were assessed, and differentiated restoration strategies were proposed according to local conditions.

Results A UAV image dataset for the Jinyun Mountain burned area was constructed. NDVI classification thresholds were determined: bare land (−1.00~0.25), burn residues (＞0.25~0.74), and healthy vegetation (＞0.74~1.00), completing preliminary land cover classification. An intelligent recognition model for vegetation recovery types was trained based on the U-Net architecture. Evaluation on the test set showed credible model performance, achieving MIoU 71.43%, OA 84.02%, mPrecision 84.21%, mRecall 82.30%, and Macro-F1 83.11%, enabling high-precision automated recognition of healthy vegetation types. Interpretation of land cover types revealed a heterogeneous recovery pattern in the study area: ferns and bamboo were the dominant pioneer species, with coverage reaching 45.04% and 16.09%, respectively, while tree coverage lagged behind at only 6.39%. Spatially, valley bottoms and lower slopes with favorable moisture conditions were the main areas for vegetation regeneration, while steep slopes with poor soil and water conservation were primarily occupied by adaptive and prolific pioneer plants like ferns and bamboo. Topographic factor extraction showed that slopes were predominantly moderately steep (53.77%), followed by steep slopes (27.52%), very steep slopes (10.76%), and gentle slopes (7.93%). Aspect and slope position categories were relatively balanced, indicating significant spatial heterogeneity. Through coupled analysis of slope, aspect, and slope position, a total of 38 types of geographic restoration units were delineated. Flow accumulation analysis further revealed differences in hydrological conditions: the top 5% high-flow accumulation zones (FA≥12602.40), mainly located in gullies and main confluence channels, exhibited strong water collection capacity; the bottom 5% low-flow accumulation zones (FA≤126.54), mainly on ridges and slopes, showed weak water collection capacity and high soil erosion risk. Integrating differences in land cover characteristics and vegetation rehabilitation conditions, an integrated ecological restoration technical pathway of “high-precision data acquisition—multi-source information intelligent interpretation—differentiated restoration strategy formulation” was established, providing a scientific basis for proposing differentiated restoration strategies for different zones within the Jinyun Mountain burned area.

Conclusion The technical pathway constructed in this study features precisional, automation, and operability. It supports the formulation of differentiated restoration strategies according to local conditions, can guide the scientific and orderly implementation of ecological restoration work, and provides a reference for the application of digital technologies in landscape architecture to serve national landscape conservation and ecological restoration. The trained deep learning semantic segmentation model possesses long-term dynamic monitoring capabilities, offering sustained support for the adaptive adjustment of restoration strategies over different periods. However, limitations exist: the current model can only identify trees, bamboo, and ferns, with limited ability to recognize shrub-layer species that may appear in mid-successional stages. Future work should expand the dataset to include more vegetation types to enhance the model's universality and generalization capacity.