Safflower harvesting has suffered from the unstable coupling between the picking head and the canopy, due to the large inter-plant heights, canopy undulations, and uneven ground in fields. In this study, a canopy-following picking control system was developed for safflower harvesting using an improved Whale Optimization Algorithm (WOA) - tuned Proportional-Integral-Derivative (PID) controller. Among them, the proportional and integral actions acted on tracking error, while the derivative action acted on the measured output. An adaptive safflower canopy-following harvester was self-designed as a research platform. A picking-head height closed-loop control system was established with a Programmable Logic Controller (PLC) as the control core, ultrasonic ranging and displacement feedback as the sensing units, and a servo electric cylinder as the actuator. A mathematical model was also established for the actuator mechanism and performance. The equivalent armature voltage was taken as the input, while the pushrod displacement of the servo electric cylinder was the output. A nominal linear model was adopted to characterize the dominant dynamic behavior of the servo motor and transmission mechanism. The Integral of Time-weighted Absolute Error (ITAE) was selected as the optimization objective. The improved Whale Optimization Algorithm (WOA) was used for offline parameter tuning. Three strategies were introduced to enhance search performance: Logistic chaotic mapping initialization to improve population diversity, a three-layer cooperative population search mechanism to balance global and local exploitation, and Lévy-flight perturbation to escape from local optima. A conventional PID and a Genetic-tuned PID controller were used as the benchmark. Step-response simulations were conducted under three typical canopy height differences of 55, 210, and 310 mm. The improved WOA-tuned PID controller achieved rise times of 0.220, 0.240, and 0.280 s, settling times of 0.300, 0.330, and 0.370 s, overshoots of 0.04%, 0.11%, and 0.18%, and steady-state errors of 0.019, 0.218, and 0.505 mm, respectively. Compared with the conventional PID and the Genetic-tuned PID controller, there was a faster dynamic response, lower overshoot, and higher steady-state accuracy under all three step conditions. Field experiments were further conducted under continuous canopy-following operation. The improved WOA-tuned PID controller shared smoother peak-to-valley transitions, smaller lag in rapidly varying sections, and faster stabilization after canopy fluctuations at a forward speed of 0.25 m/s and a tracking distance of 10 m. Its rise time and settling time reached 0.270 and 0.360 s, respectively, with the steady-state error of 0.463 mm. There was a decrease by 59.5% and 51.9%, compared with the conventional PID and the Genetic-tuned PID controller, respectively. The ITAE was reduced by 66.7% and 32.1%, respectively. The response speed, steady-state accuracy, and continuous tracking stability were effectively improved for safflower canopy-following picking. The finding can provide a strong reference for safflower canopy-following control and parameter optimization, particularly with irregular canopies.