Abstract: Unmanned aerial vehicle (UAV) remote sensing technology provides critical technical support for the precise localization, individual plant counting, survival rate estimation, and continuous growth dynamics monitoring of alfalfa (Medicago sativa L.). However, detecting dense and phenomenologically tiny objects within high-resolution UAV imagery faces severe methodological challenges. These primarily include the enormous computational overhead required for large-format image inference and the high rate of missed detections caused by severe canopy overlapping during the mid-to-late growth stages. To systematically address these bottlenecks, this study proposes a lightweight deep learning network model, designated as SFNA-YOLO (slim-faster nwd-augmented YOLO), specifically tailored for individual alfalfa localization and spatiotemporal growth dynamics monitoring via UAV remote sensing.Methodologically, using the YOLOv11 architecture as the baseline, the proposed network operates within the slicing-aided hyper inference (SAHI) framework to handle high-resolution image splicing. The model reconstructs the C3_Faster module by integrating partial convolution (PConv). This architectural modification achieves a profound decoupling of low-level feature extraction and eliminates the substantial computational redundancy inherent in standard convolutional operations. Furthermore, to compensate for localization inaccuracies associated with exceptionally small plant targets, the normalized wasserstein distance (NWD) metric is introduced into the regression branch. This integration optimizes the bounding box regression loss function by modeling targets as two-dimensional Gaussian distributions, effectively overcoming the critical scale sensitivity of tiny objects to the traditional intersection over union (IoU) metric. To ensure scale robustness, a multi-scale UAV data acquisition scheme was designed, encompassing multi-gradient vertical altitudes (15 m, 20 m, 30 m, and 60 m) alongside a 45° oblique photography perspective at 12 m.Comprehensive comparative experiments were conducted against existing counting paradigms, specifically the density-fitting-based CSNet and the point-supervision-based P2PNet, utilizing UAV time-series images collected across eight distinct phenological growth stages. Quantitative results demonstrate that the improved SFNA-YOLO model possesses an optimized parameter size (Params) of only 7.68 M, while the floating-point operations (GFLOPs) significantly decrease to 18.20. Regarding detection accuracy, the mean average precision (mAP50) reaches 95.40%, and the mean intersection over union (mIoU) substantially increases to 86.46%. These improvements effectively dismantle the computational and accuracy barriers of large-format image inference under complex field conditions.For spatial generalization, validation procedures were executed across three geographically and ecologically heterogeneous farmlands (Siziwang Banner, Tumd Left Banner, and Horinger County). Without any site-specific fine-tuning, the network accurately delineated the geometric boundaries and furrows of experimental plots, successfully overcoming environmental interferences such as severe shadow projection and varying surface albedos. Furthermore, compared to the CSNet which suffers from density adhesion and the P2PNet which experiences topological center shifts, SFNA-YOLO exhibits extremely high robustness in complex dynamic occlusion scenarios. For individual plant counting tasks, the coefficient of determination ( R^2 ) reaches 0.84, with the root mean square error (RMSE) and mean absolute error (MAE) decreasing to 0.55 and 0.30, respectively. The average survival rate error (ASR error) was controlled at a low level of 1.99%. Consequently, the model successfully facilitated the automated generation of two-dimensional spatial survival rate heatmaps and multi-temporal survival count reports for 127 specific alfalfa varieties.In conclusion, the proposed lightweight SFNA-YOLO model effectively resolves the difficulties of precise counting and growth dynamics monitoring caused by dense canopy overlapping. It significantly improves computational energy efficiency while demonstrating outstanding spatial generalization capabilities and strong temporal robustness, providing a reliable automated technical framework for high-throughput stress-resistant variety screening and refined field management operations.