Abstract: A generalized intelligent recognition framework was developed for county-scale, parcel-level orchard mapping to address the limitations of conventional deep learning approaches under small-sample conditions, particularly the difficulty of achieving high accuracy in spatially heterogeneous agricultural landscapes. Farmland parcels were initially delineated using a Boundary-Enhanced Segmentation Network (BsiNet) optimized through transfer learning on high-resolution GF-2 imagery, which enabled precise identification of parcel boundaries despite variations in shape, size, and background complexity. The resulting parcels were then spatially aligned with Sentinel-2A multispectral imagery to extract comprehensive spectral information, including multiple visible, near-infrared, and shortwave-infrared bands, for each parcel. This combination produced a high-quality parcel-level dataset integrating precise spatial geometry with rich spectral features, providing a solid foundation for subsequent classification and analysis. To enhance the extraction and fusion of deep spatial–spectral semantic features, a lightweight depthwise separable convolutional residual network, termed LiteTransResNet, was designed. The network incorporated multi-level spatial–spectral feature enhancement modules, residual connections, and depthwise separable convolution operations to improve representational capacity, enabling the model to capture both local spatial structures and subtle spectral variations while maintaining computational efficiency suitable for large-scale agricultural applications. To overcome the limitations of scarce labeled data, a semi-supervised learning strategy was employed to dynamically expand the labeled dataset by selecting high-confidence predictions from unlabeled parcels. This approach reduced the dependency on manual annotation and significantly enhanced the generalization capability of the model for cross-regional orchard classification tasks. The integrated framework, combining BsiNet-based parcel delineation with semi-supervised LiteTransResNet classification, was applied to plum orchards in Jiashi County, Xinjiang, China, and validated using field survey records and planting area statistics provided by local agricultural authorities. In Baren and Mixia townships, the proposed method achieved overall classification accuracies of 99.11% and 99.46%, respectively, substantially outperforming conventional deep learning methods and demonstrating the effectiveness of combining spatial–spectral semantic feature enhancement with semi-supervised learning under limited sample conditions. Furthermore, when trained using only plum samples from these two townships, the model estimated plum planting areas across twelve townships of Jiashi County with errors ranging from 2% to 9%, highlighting its robust generalization capability despite minimal training data. The framework also generated high-resolution farmland parcel data at 1-meter spatial resolution and orchard distribution maps at 10-meter resolution, effectively capturing both local spatial patterns and multispectral characteristics, which facilitated accurate county-scale orchard mapping and provided reliable inputs for precision agricultural management. The results indicated that integrating spatial–spectral semantic feature enhancement with semi-supervised learning substantially improved classification performance under small-sample conditions, while the lightweight network structure ensured efficiency and scalability for extensive mapping tasks. Overall, the study presented a reliable approach for high-precision, parcel-level orchard recognition and spatial mapping under constrained sample availability, confirming that county-wide orchard distributions could be accurately estimated using minimal labeled data. The methodology provided strong support for regional-scale orchard monitoring, crop type mapping, growth assessment, yield estimation, and precision agriculture applications. Additionally, the framework offered a transferable solution for other high-resolution, small-sample agricultural remote sensing tasks, demonstrating the potential of combining advanced deep learning architectures with semi-supervised learning strategies to achieve robust generalization and scalable performance across heterogeneous landscapes. These findings underscored the value of integrating high-resolution spatial information, multispectral features, and adaptive training strategies to enhance both the accuracy and operational efficiency of large-scale orchard classification and monitoring, providing a practical and replicable reference for future precision agriculture research and implementation.