| 中文题名： | 基于深度卷积神经网络的大尺度农作物遥感制图 |
| 姓名： | |
| 保密级别： | 公开 |
| 论文语种： | 中文 |
| 学科代码： | 081602 |
| 学科专业： | |
| 学生类型： | 硕士 |
| 学位： | 工学硕士 |
| 学位类型： | |
| 学位年度： | 2021 |
| 校区： | |
| 学院： | |
| 第一导师姓名： | |
| 第一导师单位： | |
| 提交日期： | 2021-06-08 |
| 答辩日期： | 2021-06-04 |
| 外文题名： | LARGE AREA CROP REMOTE SENSING MAPPING USING DEEP CONVOLUTIONAL NEURAL NETWORK |
| 中文关键词： | |
| 外文关键词： | Deep learning ; Deep convolutional neural network ; Spatial-temporal generalization ; Weak samples ; Large-area crop mapping |
| 中文摘要： |
准确的大尺度农作物空间分布是农作物生长监测、种植面积估算和产量预测的基础，对国家粮食安全及社会的可持续发展具有重要的研究和生产意义。传统方法分类过程耗时耗力，所提取特征难以实现跨时空迁移，泛化能力差，无法实现高效的大尺度农作物遥感制图。大数据驱动下，深度学习“端到端”的自动分类方式能够克服传统方法的不足，但存在对大量人工标记样本的高度依赖问题。
﹀
基于此，本研究采取了一种基于深度学习模型的策略，首先对比分析深度学习模型和传统分类器随机森林（Random Forest, RF）在农作物遥感分类过程中跨年际和跨地区的识别结果，验证深度学习模型时空泛化性内在机理；其次验证传统分类方法支持向量机（Support Vector Machines, SVM）生产出的大量较高精度的标记样本（记为“弱样本”）进行深度学习训练的适用性；最后利用弱样本的深度卷积神经网络（Deep Convolutional Neural Network, DCNN）模型对跨省级数据集进行大尺度农作物遥感自动化分类，得到以下结论： （1）DCNN模型具有良好的时空泛化能力。利用DCNN模型和RF模型针对空间泛化集、时间泛化集和时空泛化集3个不同时空数据集进行冬小麦提取，对分类结果进行整体定量对比和子区目视比较。在基于混淆矩阵的精度评价统计结果中，DCNN模型的3个时空泛化集中OA和冬小麦F1分数均高于0.90，对于OA，比RF模型结果精度分别增加了0.11、0.14和0.19；对于F1分数，分别提升了0.14、0.17和0.29，具有更好的分类精度；DCNN模型在3个不同时空泛化集之间的精度标准差低于RF模型，具有更强的稳定性，能够避免基于像元分类浅层模型的“椒盐噪声现象”，具有捕捉语义、细节特征等特征的能力，这些特征能表征冬小麦作物的类别归属特征和形状、位置、边界等细节特征，使模型在不同时空测试集中拥有稳定的强泛化能力。 （2）DCNN模型对弱样本具有一定的适用性。基于传统分类方法得到作物分类结果作为标签样本训练深度学习模型能够得到较高的识别精度，在不同地形地貌的农业景观下保持良好的分类性能。采用可靠精度的SVM分类结果作为弱样本对DCNN模型进行训练，预测结果OA达到0.90，其中水稻和玉米的F1分数分别为0.81和0.90，与SVM结果精度一致性达到0.90。通过子区分析，DCNN模型在平原和山地两种地貌类型中均表现出较好的分类效果，结果OA中值均大于0.93，在一定程度上克服了研究区地形带来的影响。通过噪声实验证明了当模型在5倍样本噪声内，即样本最大误差面积比例不超过0.36时，模型具有一定的鲁棒性，结果能保持在可靠的精度范围内。该方法弥补了深度学习模型对大量人工标记样本高度依赖的缺陷。 （3）基于弱样本的DCNN模型能够自动化生产较高精度的省级大尺度农作物空间分布产品。基于实现大尺度农作物遥感制图的目标，在本研究对DCNN模型时空泛化性研究和DCNN模型弱样本的适用性研究的基础下，利用辽宁省弱样本训练得到的DCNN模型对同年吉林省全省水稻和玉米进行分类。从整体结果精度上看，OA为0.84，水稻和玉米的分类结果保持较高精度，F1分数分别为0.82和0.88，表明该策略基本实现了省级大尺度农作物制图，验证了DCNN模型具备较强的泛化能力，表明了利用弱样本训练出来的深度学习模型具有可行性。基于山区和平原区的格网OA、PA、UA和F1分数指标的精度的可视化结果表明了DCNN模型的泛化分类性能对地形因素具有一定的鲁棒性。通过与SVM结果比较表明了DCNN模型优于基于人工干预的“特定区域，特定样本”分类方法。因此说明了利用弱样本训练得到的DCNN模型不仅解决了样本数据的标记问题，而且具有良好的泛化能力，在一定程度上可以实现高效的大尺度农作物自动化制图。 |
| 外文摘要： |
Accurate spatial distribution of large area crops is the basis of crop growth monitoring, planting area estimation and yield prediction, which has important research and production significance for national food security and sustainable development of society. The classification process of traditional methods is time-consuming and labor-intensive, and the extracted features lacked the flexibility across time and space with the poor generalization ability, which makes it impossible to achieve efficient large area remote sensing mapping of crops. Driven by big data, the "end-to-end" automatic classification method of deep learning (DL) can overcome the shortcomings of traditional methods, however, it has the problem of high dependence on a large number of manually labeled samples.
﹀
Based on this, a strategy based on deep learning model was introduced in this study. the cross-annual and cross-regional classification results of DCNN model and traditional classifier Random Forest (RF) in the process of crop remote sensing classification were compared and analyzed to verify the inherent mechanism of spatial-temporal generalization of deep learning model. Secondly, the applicability of a large number of high-precision labeled samples (referred to as "weak samples") produced by traditional classification method Support Vector Machines (SVM) for deep learning training was verified. Finally, a Deep Convolutional Neural Network (DCNN) model with weak samples was used to automate large area crop remote sensing classification of trans-provincial datasets. And the following conclusions were obtained: (1) DCNN model has good spatial-temporal generalization ability. The DCNN model and RF model trained on the same dataset were used to extract winter wheat for three datasets, which are the spatial generalization dataset, the temporal generalization dataset and the spatial-temporal generalization dataset, respectively. Then, the classification results were compared quantitatively as a whole and the subregions were compared in visual comparison. In the statistical results of accuracy evaluation based on confusion matrix, the OA and F1 score of winter wheat in the three spatial-temporal generalization datasets based on DCNN model are all higher than 0.90, which is more accurate than the results of RF model. For OA, the accuracy of DCNN model is 0.11, 0.14 and 0.19 higher than that of RF model. For the F1 score, it is improved by 0.14, 0.17 and 0.29, respectively, with better classification accuracy. Moreover, the accuracy standard deviation of DCNN model is lower than that of RF model in three different spatial-temporal generalization datasets, so it has stronger stability. It can avoid the "salt and pepper noise phenomenon" based on the pixel classification shallow model, and has the ability to capture the semantic and detailed features, which can represent the classification characteristics of winter wheat and the detailed features such as shape, position, boundary, etc., so that the model has a stable and perfect generalization ability in different time and space testing datasets. (2) The DCNN model has certain applicability to weak samples. Crop classification results obtained based on traditional classification methods can be used as label samples to train the deep learning model, which can obtain high recognition accuracy and maintain good classification performance in agricultural landscapes with different topographies and landforms. The classification results of SVM with reliable accuracy were used as weak samples to train the DCNN model. As a result, the OA of the testing results reached 0.90, and the F1 score of rice and maize were 0.81 and 0.90, respectively, with the accuracy consistency between the testing results and SVM results being 0.90. Through the subregion analysis, the DCNN model shows a good classification effect in both plain and mountain area. And results showed that the median of OA was greater than 0.93, which illustrated the model can overcome the influence of topography in the study area to a certain extent. Noise experiments show that when the model is within 5 times noise of the sample, that is, the proportion of the maximum error area of the sample is not more than 0.36, the model has certain robustness, and the results can be kept within a reliable accuracy range. This method makes up for the defect of deep learning model which is highly dependent on a large number of manually labeled samples. (3) The DCNN model based on weak samples can produce high-precision provincial large-area agricultural spatial distribution mapping automatically. In order to achieve the goal of large-area crop remote sensing mapping, based on the study on spatial-temporal generalization of DCNN model and the applicability of weak samples of DCNN model in this study, the DCNN model trained with weak samples in Liaoning Province was used to classify rice and corn in Jilin Province in the same year. Looking from the overall results accuracy, OA reached 0.84, and the classification of rice and corn remained relatively high precision, whose F1 score were 0.82 and 0.88, respectively, showing that the proposed strategy can achieve provincial large-area crop mapping basically, and further verifying the DCNN model has strong ability of generalization and the flexibility of the deep learning model trained with weak samples. Visualization of the accuracy of OA, PA, UA and F1 score of each grid in mountainous and plain areas showed that the generalized classification performance of DCNN model was robust to topographic factors to some extent. Moreover, the comparison between the results of DCNN and SVM showed that the DCNN model was superior to the "specific region, specific sample" classification method based on manual intervention. Therefore, it showed that the DCNN model obtained by using weak sample training not only solved the labeling problem of sample data, but also had good generalization ability, and can achieve efficient large-area automatic mapping of crops to a certain extent. |
| 参考文献总数： | 100 |
| 馆藏号： | 硕081602/21008 |
| 开放日期： | 2022-06-08 |
谷歌
