摘要:
土地覆盖类型检测对地表特征理解、生态环境监测和自然资源管理具有重要意义。基于哨兵1号（Sentinel-1）多时相相干性矩阵，搭建长短期记忆网络（long short-term memory,LSTM）土地覆盖分类模型，旨在评估其在土地覆盖分类中的效能，并探讨不同分类方法对分类结果的影响，并以中国辽宁省阜新市彰武县为研究区对4种不同土地覆盖分类方法进行验证。实验结果显示，基于Sentinel-1相干性矩阵的LSTM分类模型方法在3个区域的分类精度分别高达93.7%、89.9%和87.6%，具有较高的分类精度和鲁棒性，比基于Sentinel-1相干性向量的支持向量机（support vector machine,SVM）分类模型、基于Sentinel-1相干性矩阵的卷积神经网络分类模型以及基于哨兵2号（Sentinel-2）多光谱影像的SVM分类模型的方法分类精度高约1%～10%。此外，基于Sentinel-1相干性矩阵的LSTM分类模型方法还在不同子研究区之间的进行了迁移学习实验，结果显示，该模型在子区域迁移时的泛化能力仍需增强。
Abstract:
Objectives: Land cover classification is crucial for understanding surface features, monitoring ecological environments, and managing natural resources. The objective of this research is to construct a long short-term memory (LSTM) model for land cover classification using Sentinel-1 multi-temporal coherence matrices and to evaluate its effectiveness by comparing it against other three classification methods. Methods: Multi-temporal coherence matrices from Sentinel-1 synthetic aperture radar images were used to train an LSTM neural network for land cover classification. The Sentinel-1 coherence matrix was firstly transformed to sequences according to the temporal baseline intervals and then input into the LSTM model to predict the land cover types. This LSTM-based method was then compared against other three methods: (1) support vector machine (SVM) and (2) convolutional neural network models taking the Sentinel-1 coherence vectors and matrices as input, and (3) SVM model with the input of Sentinel-2 multispectral features. Transfer learning experiments were also conducted to assess the generalization capability of the LSTMbased method. Results: The study area was selected in Zhangwu County, Fuxin City, Liaoning Province, China, where three test regions are chosen to conduct the land cover classification by using the four different methods, respectively. Experimental results demonstrated that the LSTM model with the Sentinel-1 coherence matrices achieved classification accuracies of 93.7%, 89.9% and 87.6% in the three test regions, respectively, with an improvement in classification accuracy by approximately 1% to 10% compared to other methods. However, the results of the transfer learning experiments reveal that the generalization capability of the LSTM model is not desirable and needs further enhancement. When trained solely on region 1 and tested on regions 2 and 3, classification accuracies dropped to 45.7% and 47.4%. To enhance classification performance and model generalizability, a comprehensive sampling strategy was implemented by integrating data from all three regions for unified training. This approach achieved regional classification accuracies of 92.8%, 87.8%, and 86.4%, effectively mitigating sample distribution bias and significantly improving cross-region adaptability. Conclusions: The LSTM model utilizing Sentinel-1 coherence matrices as input features shows promising performance in land cover classification, particularly in areas affected by cloud cover. However, transferability of the model has high demand for the training samples, which can inhibit its applicability to wider regions. Future research is needed to optimize transfer learning strategies to enhance the model's adaptability and classification accuracy across different experimental regions.
下载:
