Aiming to address the issues of low efficiency and insufficient automation in traditional soil available nitrogen, phosphorus and potassium (NPK) content detection, as well as the limited path planning and tracking performance of vehicle-mounted detection equipment in complex farmland environments, an autonomous detection system for soil available NPK in field empowered by an unmanned operation platform was established. It integrated Beidou positioning and a four-wheel drive steering platform, optimized sensor detection models, designed a hierarchical-adaptive traversal path planning method for fertilizer-measuring points and a path tracking control algorithm based on genetic algorithm (GA)-optimized linear quadratic regulator (LQR), and built an autonomous fertilizer detection system with cloud-edge architecture. For in-situ, rapid detection of soil fertility indicators, random forest (RF) prediction mode was selected as the optimal algorithm after comparing four machine learning models. Its relative prediction errors for available N, P and K were below 16.44%, 19.26% and 13.91%, respectively, with an average 49.46% accuracy improvement over the direct measurement accuracy of sensors without prior modeling. To accurately characterize the spatial distribution of field-scale fertility indicators, the optimal detection point spacing was determined by quantifying the Christiansen uniformity coefficient (CUC) across different sampling densities;three continuous traversal path planning schemes—comb path (CP), comb cross-row path (CCP), and comb fifth-order Bézier path (CFBP)—were designed for flat obstacle-free, large-obstacle, and small-obstacle farmland environments, respectively. Simulation tests showed that the GA-LQR controller improved the platform’s path tracking accuracy by an average of 24.32%, with maximum absolute lateral deviation and heading angle of 1.71cm and 1.69°, respectively. Field tests demonstrated that under straight path tracking, the GA-LQR algorithm reduced the maximum absolute lateral deviation, maximum absolute heading angle, average absolute parking error at detection points, and total detection time by 15.62%, 19.59%, 20.79%, and 13.43%, respectively, compared with the conventional LQR algorithm;under curved path tracking, the corresponding indicators were decreased by 19.28%, 27.34%, 22.94%, and 15.76%. Additionally, the optimized algorithm shortened the detection time per hectare by 927s, reduced the single-point detection process by 6.28s, and improved the positioning and detection accuracy by an average of 14.26%. The research result can provide a reference for the autonomy, informatization, and intelligentization of soil available NPK content information acquisition.