A Rocky Desertification Land Extraction Method Based on Spectral-Texture-Scattering-Terrain Multisource Features With Time Series

In karst regions, rocky desertification (RD) represents an extreme form of land degradation, posing significant threats to both the ecological environment and socio-economic development. Extracting the spatial distribution of rocky desertification land (Roc) is crucial for RD management, resource allocation, and sustainable development. However, current research primarily focuses on RD degree classification, with lack of studies distinguishing Roc from bare soil land (Bar) and other land types, neglecting the dynamic degradation characteristics of Roc in single-temporal data, and facing challenges in effectively integrating high-dimensional nonlinear features with machine learning (ML) methods, leading to confusion between Roc and Bar and inaccurate Roc identification. To address these challenges, this study first constructs a time-series-based Spectral-Texture-Scattering-Terrain (STST) feature set and applies multi-scale segmentation (MS) processing. Subsequently, by incorporating an improved spatial-channel attention mechanism and a gating mechanism-based classification module, a Multi-Scale Dilated Novel Attention 1-Dimensional Convolutional Neural Network (MDNA-1DCNN) is designed for object-oriented classification (OOC) in Ziyun County, Guizhou Province, and Longlin County, Guangxi Province, China. This method does not require cutting the images into small patches for network input, enabling more efficient large-scale Roc extraction. Experimental results show that MDNA-1DCNN achieves an overall accuracy (OA) of 97.67% and a Kappa coefficient of 97.07% in multi-class tasks, with F1-score (F1) and Intersection Over Union (IoU) for Roc extraction reaching 97.02% and 94.21%, respectively, outperforming other advanced deep learning (DL) networks and ML methods. Moreover, the STST feature set, which encompasses multiple attribute features from various perspectives, is more effective in improving Roc extraction accuracy compared to using only 1-3 of these attributes. This study provides effective technical support and reference for land monitoring, RD prevention, and ecological restoration in karst vulnerable regions.