This study focuses on developing a strategy for constructing an adaptive training space for neural network-based forest canopy parameter inversion, with the goal of enhancing LAI estimation accuracy by optimizing the training space sampling process. To address the accuracy limitations of traditional methods in regions with scarce samples, this approach introduces an adaptive sampling mechanism that aims to create a more regionally representative training sample set. By constructing the regionalized PROSAIL model parameter probability distributions and integrating them into the machine learning framework as prior knowledge, high-precision LAI prediction for independent validation datasets was ultimately achieved. Experimental results demonstrate that the proposed LAI inversion method, which leverages an adaptively constructed training space, achieves high estimation accuracy with results closely matching ground-truth measurements. Specifically, the validation results for forest ecosystems show that the root-mean-square error of the method is reduced to 57.99%. This study reveals that the method not only improves the estimation accuracy, but also successfully reveals the characteristics of joint probability distribution of PROSAIL model parameters under complex environmental conditions. This innovative parameter optimization framework provides a new idea for accurate remote sensing inversion, and also shows good generalization ability, which provides a solid theoretical foundation for wider ecosystem monitoring and vegetation parameter estimation in the future.