Abstract: Straw row-consolidation has one of the most representative patterns during conservation tillage in the black soil region of Northeast China. Straw-covered rows and clean seedling bands are alternated to improve soil conservation and seedbed quality. However, conventional full-coverage pesticide application can lead to chemical waste and environmental contamination because spray droplets are intercepted by straw residues rather than deposited on target crop zones. Therefore, it is often required to accurately and real-time extract the crop seedling band centerline for intelligent inter-row spraying and pesticide application in precision agriculture. Challenges also remain on complex straw texture interference, weak grayscale contrast between soil and straw, and the real-time constraints of agricultural machinery. In this study, an adaptive multi-threshold Otsu segmentation was proposed using improved swarm intelligence optimization. An inter-row spraying was also developed using crop seedling band recognition. Field experiments and demonstrations were conducted at seven conservation tillage sites in Jilin Province, Northeast China. Straw row-consolidation planting structure was alternated between 70 cm straw rows and 60 cm seedling band rows. An image acquisition and processing platform was established for accurate recognition under real conditions. An industrial camera (2K resolution with 2.8 mm focal length) was mounted at the front of a plant protection tractor with a height of 3 m. Comparative tests were performed under different camera depression angles. A 20° oblique depression angle was selected as the optimal acquisition. An effective acquisition width of 9 m was reserved 3 s for image processing, enabling synchronized operation when the tractor moved at 10 km/h. The captured images were transmitted into an embedded machine vision processor (NVIDIA Jetson Orin Nano 4GB, Ubuntu 20.04 LTS) for real-time detection of seedling band centerline. The coordinate information was extracted and then delivered into an STM32 microcontroller in the spraying execution module. Furthermore, color images were first converted into greyscale using YUV luminance channel. Computational efficiency was improved in the vision. Otsu thresholding was applied for binarization. A denoising strategy was used to suppress irregular noise caused by fragmented straw residues, according to component area filtering and morphological opening. A hybrid swarm intelligence optimization, named SADE-WOA, was proposed to integrate differential evolution (DE), simulated annealing (SA), and the whale optimization algorithm (WOA). There were low computational complexity and less susceptibility to local optima, compared with the conventional Otsu multi-threshold segmentation. Specifically, the DE crossover mechanism was embedded into the position update stage of WOA, and vector perturbation was incorporated to enhance population diversity and suppress premature convergence. Meanwhile, SA was used to strengthen local refinement and escaping. Multi-threshold segmentation task was reformulated as a global optimization in a high-dimensional threshold combination space. Adaptive threshold selection was realized for the images from the complex fields. Morphological operations were employed to enhance region connectivity after segmentation. Discontinuities and adhesion were also reduced among seedling band regions. A watershed was applied to separate the regions of interest using seedling bands. Canny edge detection was then used to extract left and right boundary points. The boundary coordinates were extracted to generate the midpoint sets. The seedling band centerline was finally fitted using the least squares method. Comparative experiments were conducted to evaluate the performance, convergence stability, and computational efficiency of the SADE-WOA segmentation. Results show that the SADE-WOA was accurately distinguished soil type to preserve straw texture information, particularly for the maximum between-class variance and stable convergence. The average computation time per image was only one-tenth that of the grey wolf optimization (GWO) and one-ninth that of the conventional differential evolution. Centerline detection accuracy was further evaluated using mean absolute error (MEA), root mean square error (RMSE) and mean relative error (MRE). The yaw angle error of seedling band centerline ranged from 0.34° to 0.64°, and the maximum relative error was below 0.2%, indicating the low cumulative deviation and the high angular accuracy in fields. A modular precision spraying device was designed to integrate camera sensing, embedded vision computing, stepper-motor-driven nozzle positioning, and fuzzy PID flow control. A linear sliding module drove the nozzle to detect seedling band position for inter-row spraying. A fuzzy PID controller dynamically adjusted the spray flow rate in real time. Droplet drift was avoided to improve deposition uniformity. Field spraying trials were conducted at the Changchun Agricultural Machinery Research Institute conservation demonstration base on October 21, 2025. Three test plots of 200 m length were selected during experiments. The tractor operated at a constant speed of 10 km/h. A pesticide saving rate of 54.3% was achieved in the improved system, compared with conventional full-coverage spraying. Droplet coverage was maintained between 27.5% and 32.1%. The uniformity of pesticide deposition was also improved after SADE-WOA-driven segmentation and centerline extraction with the modular inter-row spraying. The findings can also provide an effective solution for precision pesticide application under straw row-consolidation conservation tillage, thus contributing to pesticide utilization efficiency with minimal pollution.