Aiming at the problem of single-machine operation task planning in adjacent multiple fields in the same area, according to the practical limits such as battery life and load-bearing, a path planning algorithm for an autonomous agricultural rotary-wings UAV with multi-flight operations in multiple fields with obstacles was proposed and realized considering the on-demand multiple returning supply of consumables such as seeds, fertilizer, chemicals, oil and electricity. On the basis of giving the operation path planning optimization target model, a two-step optimization method of “first path, and then flights” was proposed. Firstly, the initial coverage working path was planned with the shortest path as the optimization goal, and then the flights planning was carried out considering the replenishment and sequels after the returns of the consumables on demand. According to the two kinds of spraying and spreading modes, including the variable mode and constant mode, two orientation arrangements, route scheduling principle, replenishment mode and sequel mode were proposed respectively. The initial operation paths under four different headings and route scheduling strategies was planned and obtained. Then the flights arrangement solution under four different combinations of replenishment optimization strategies was further given. The simulation experiments on hypothetical plots and real plots showed that the proposed algorithm with the four optimization strategies of the initial operation paths run very stable and efficient. The first stage algorithm took time from 78ms to 105161ms, saving the transfer path from 10.25% to 33.81% when compared the better results with the relatively poor results. The time consumption of the second stage algorithm was increased slightly from 21170ms to 106393ms, saving the replenishment times from 15 to 10 in a test calculation and the transfer path from 14.00% to 26.27% when compared the better results with the relatively poor results. In general, the combine replenishment strategy of “all reset” and “go-on with new order” got better performance than others.