Accelerating forest stand selection for subsidization using neural networks
Accelerating forest stand selection for subsidization using neural networks
复制标题
DOI：
10.1016/j.compag.2022.107595
复制
发表时间：
2023-02
期刊：
影响因子：
--
通讯作者：
Kai Moriguchi;Hiroaki Shirasawa;K. Aruga
中科院分区：
文献类型：
--
作者：
Kai Moriguchi;Hiroaki Shirasawa;K. Aruga
林分选择补贴方法需要对每一个林分的最佳采伐计划进行迭代确定，迭代计算所需的时间以及在重点区域存在大量的林分，阻碍了林分选择方法在实践中的应用。在这项研究中，我们开发了两种方法，以减少计算成本的林分选择使用神经网络。一种方法使用神经网络近似林分选择所需的指标，从而降低计算成本。另一种方法减少了迭代优化过程中的计算成本，并间接预测所需的指标。第一种方法即使在不平衡的训练数据下也能对必要的指数进行可接受的预测。虽然使用第二种方法生成的神经网络具有合理的精度，但由于间接预测，预测指数具有不可消除的误差。第一种方法使用20，000个调整和验证数据生成具有合理精度的神经网络，当应用于日本长野县时，将计算时间从61.09 h减少到7.62 h。
A method of forest stand selection for subsidization requires iterative identification of optimal harvesting schedules for each stand. The time required for the iterative calculation and the existence of numerous stands in the focused regions prevent the application of the stand selection method in practice. In this study, we developed two methods to reduce the calculation cost of stand selection using neural networks. One method approximates the necessary indices for stand selection using neural networks, thereby reducing the calculation cost. The other method reduces the calculation cost during the iterative optimization processes and indirectly predicts the necessary indices. Acceptable predictions of the necessary indices were made by the first method even with imbalanced training data. Although neural networks with reasonable accuracy were generated using the second method, the predicted indices had unignorable errors owing to indirect predictions. The first method generated neural networks with reasonable accuracy using 20,000 tuning and validation data, which reduced the computation time from 61.09 h to 7.62 h when applied in Nagano Prefecture, Japan.