In gradient computations of the variational data assimilation (VDA) by the adjoint method, in order to overcome a lot of shortcomings such as low accuracy, difficult implementation, and great complexity, etc., a novel data assimilation method is proposed based on the dual-number theory. The important advantages are that the coding of adjoint models and reverse integrations are not necessary any more, and the values of cost functional and its corresponding gradient vectors can be attained simultaneously only by one forward computation in dual-number space. Furthermore, the accuracy of gradient can be close to the computer machine precision without other error sources. The paper is organised as follows. Firstly, the dual-number theory and algorithm rules are introduced. Then, the issues of gradient analysis and computation in VDA are transformed into the processes of calculating the cost functional numerically in dual-number space, and the gradient vectors can be obtained at the same time in an easy, efficient and accurate way. Secondly, the new algorithm for data assimilation in nonlinear physical systems is developed by combining accurate gradient information from the dual-number method with classical optimization algorithm. Thirdly, numerical experiments on sensitivity analysis for an ENSO nonlinear air-sea coupled oscillator are implemented, and the results are presented to demonstrate the important advantages of the dual-number method in the calculation of derivative information. Finally, numerical simulations for data assimilation are carried out respectively for the typical Lorenz 63 chaotic systems, the specific humidity evolving equation with physical “on-off” process at a single grid point, and a parabolic partial differential equation. Some conclusions can be drawn from the numerical experiments. The newly proposed method may be suited to many kinds of optimization problems with ordinary or partial differential equations as constraints, such as data assimilation, parameter estimation, inverse problems, sensitivity analysis etc. Results show that the new method can reconstruct the initial conditions or parameters of a nonlinear dynamical system very conveniently and accurately. Its another advantage is being very easy to implement with a high accuracy in gradient computation, so it is robust in the process of numerical optimization. The estimated initial states or parameters are convergent to real value in the cost of no more computations, when there are noises in the observations. But many tests are still needed to demonstrate the validity and advantages of the new data assimilation method, especially in more complex and realistic numerical prediction models of atmosphere and ocean.