森林作为陆地最重要的生态系统，在生态平衡、经济建设以及水源涵养等方面发挥了巨大的价值。因此，进行森林资源的管理及监测具有重要意义。传统的森林资源调查成本高、难度大、效率低，而遥感技术大大提高了森林资源调查的效率，为森林类型识别提供了新的机遇。但遥感分类受到数据源、地物特征以及分类方法等因素影响，因此需要在数据源选取、特征提取与优选以及分类模型等方面进行探究。

       本文将黑龙江省孟家岗林场作为主要研究区，以多时相Sentinel-1 SAR、Sentinel-2以及GF-1影像为遥感数据源，综合数字高程模型（Digital Elevation Model，DEM）、CCD影像、二类调查小班数据以及实地调查数据，构建林场内的分类体系并结合林分类型物候特性，提取多时相光谱特征、植被指数、雷达特征、纹理特征以及地形因子，分析并优选有利于区分各类别的特征，构建多种分类特征组合，探究多源多时相数据下森林林分类型的分类效果以及不同特征优选算法下优选变量对分类精度的影响。同时，进一步探究深度学习方法（U-Net、SegNet、DeepLab V3+）在森林林分类型中的分类效果及适用性。主要研究内容和结论如下：

（1）多源遥感影像的林分类型分类。通过分别对多时相Sentinel-1/2影像和GF-1影像的所有特征，采用随机森林方法进行分类实验，发现多时相Sentinel-1/2影像所有特征的分类精度为82.88%，比多时相GF-1影像所有特征精度高2.05%，说明丰富的光谱信息更加利于影像分类。同时，综合Sentinel-1 SAR影像、Sentinel-2影像以及GF-1影像的所有特征进行分类时，其精度达到83.33%，比采用单一数据源精度高。因此，综合多源影像数据有利于提高分类精度。

（2）基于不同特征优选算法的多特征影像林分类型分类。针对多特征影像中可能存在冗余变量，对Sentinel-1/2影像所有特征、GF-1影像所有特征以及综合Sentinel-1/2和GF-1影像所有特征三种方案，分别采用VSURF、Boruta、RFE以及varSelRF四种特征优选算法进行变量筛选并采用随机森林方法分类，结果表明varSelRF方法优选变量的效果最好，能够减少冗余变量并提高模型精度。通过变量优选后三种方案分类精度分别提高了0.32%、0.42%以及0.92%，表明特征优选能够避免变量冗余，从而提高模型效率。

（3）基于U-Net模型的不同空间分辨率影像林分类型分类。对Sentinel-2影像光谱特征+DEM和GF-1影像光谱特征+DEM两种方案分别采用最大似然法、支持向量机、决策树、随机森林以及U-Net模型进行分类，结果表明两种方案中U-Net模型的精度明显优于其他分类方法；支持向量机次之，决策树精度最低。同时，基于Sentinel-2影像的U-Net模型精度比GF-1影像高4.5%，说明U-Net模型能够学习到影像丰富的多波段特征信息，进而提高精度。

（4）基于深度学习方法的多源影像林分类型分类。通过结合多时相Sentinel-2影像和GF-1影像的光谱特征以及DEM数据，分别采用三种深度学习方法U-Net、SegNet和DeepLab V3+模型进行分类，并与传统的机器学习方法进行对比，结果表明三种深度学习方法的分类精度均比传统的机器学习方法高，其中U-Net模型的总体精度最高为86.08%；之后依次是DeepLab V3+、SegNet模型；传统的机器学习方法中最大似然法精度较高为80.55%，比随机森林高3.37%。深度学习模型能够自动学习并挖掘影像深层特征信息，减少椒盐噪声，有效地提高分类精度，为后续森林林分类型分类研究提供参考。

As the most important terrestrial ecosystem, forests have played a huge role in ecological balance, economic construction and water conservation. Therefore, the management and monitoring of forest resources is of great significance. The traditional forest resources survey is expensive, difficult and low in efficiency, while remote sensing technology has greatly improved the efficiency of forest resources survey and provided new opportunities for forest type identification. However, remote sensing classification is affected by factors such as data sources, feature features, and classification methods. Therefore, it is necessary to explore data source selection, feature extraction and optimization, and classification models.

This paper takes Mengjiagang Forest Farm in Heilongjiang Province as the main research area, and uses multi-temporal Sentinel-1 SAR, Sentinel-2 and GF-1 images as remote sensing data sources, integrates digital elevation model, CCD image, forest management inventory data and field survey data, constructs the classification system in the forest farm, and extracts multi temporal spectral features, vegetation index, radar features, texture features and terrain factors combined with the phenological characteristics of forest types. Then analyze and optimize the features that are conducive to distinguishing various categories, construct a variety of classification feature combinations, and explore the classification effect of forest stand types under multi-source and multi temporal data and the influence of optimization variables under different feature optimization algorithms on classification accuracy. At the same time, further explore the classification effect and applicability of deep learning methods (U-Net, SegNet, DeepLab V3+) in forest stand types. The main research contents and conclusions are as follows:

(1) Classification of stand types from multi-source remote sensing images. By classifying all the features of the multi-temporal Sentinel-1/2 images and GF-1 images separately, using the random forest method to conduct classification experiments, it is found that the time accuracy of synthesizing all features of sentinel-1/2 image is the highest, which is 82.88%; When integrating all features of GF-1 image, the highest accuracy is 80.83%. The user accuracy and producer accuracy of each category reach the maximum, and the misclassification and misclassification in the confusion matrix are reduced. It can be seen that the Sentinel-1/2 image has higher accuracy, which is 2.05% higher than that of the GF-1 image, indicating that the rich spectral information is more conducive to image classification. Combining all the features of Sentinel-1 SAR image, Sentinel-2 image and GF-1 image, the classification accuracy reaches 83.33%, which is higher than that of using a single data source. Therefore, integrating multi-source image data is beneficial to improve the classification accuracy.

(2) In view of the possible redundant variables in multi feature images, four feature optimization algorithms of VSURF, Boruta, RFE and varselRF are used for variable screening and random forest method classification for all the features of sentinel-1/2 image, all the features of GF-1 image and all the features of sentinel-1/2 and GF-1 image. The results show that varSelRF method has the best effect on optimizing variables, which can reduce redundant variables and irrelevant variables and improve the accuracy of the model. Through variable optimization, the classification accuracy of the latter three schemes is improved by 0.32%, 0.42% and 0.92% respectively, indicating that feature optimization can avoid variable redundancy and improve the efficiency of the model.

(3) Classification of image forest stand types with different spatial resolutions based on U-Net model. The maximum likelihood method, support vector machine, decision tree, random forest and U-Net model are used to classify the two schemes of Sentinel-2 image spectral feature + DEM and GF-1 image spectral feature + DEM respectively. The results show that the accuracy of the U-Net model in the two schemes is significantly better than other classification methods; the support vector machine is the second, and the decision tree Kyoto is the lowest. At the same time, the accuracy of the U-Net model based on Sentinel-2 image is 4.5% higher than that of GF-1 image, indicating that the U-Net model can learn the rich multi-band features of the image, thereby improving the accuracy.

(4) Multi-source image forest stand types classification based on deep learning method. By combining the spectral characteristics of multi temporal sentinel-2 image and GF-1 image and DEM data, three deep learning methods U-Net, SegNet and DeepLab V3+ model are used to classify forest stand types, and compared with maximum likelihood method and random forest. The results show that the three deep learning methods have higher classification accuracy than the traditional machine learning. Among them, the accuracy of U-Net model is the highest, with an overall accuracy of 86.08% and a kappa coefficient of 0.8163, followed by DeepLab V3+ and SegNet models. Among the traditional machine learning methods, the accuracy of maximum likelihood method is 80.55%, which is 3.37% higher than that of random forest classification. The deep learning model can automatically learn and mine the deep feature information of the image, reduce the salt and pepper noise, effectively improve the classification accuracy, and provide a reference for the subsequent research on the classification of forest stand types.

[1]何兴元, 任春颖, 陈琳, 等. 森林生态系统遥感监测技术研究进展[J]. 地理科学, 2018, 38(07): 997-1011.

[2]杨超, 邬国锋, 李清泉, 等. 植被遥感分类方法研究进展[J]. 地理与地理信息科学, 2018, 34(04): 24-32.

[3]张颖, 李晓格, 温亚利. 碳达峰碳中和背景下中国森林碳汇潜力分析研究[J]. 北京林业大学学报, 2022, 44(01): 38-47.

[4]王怀警, 谭炳香, 王晓慧, 等. 多分类器组合森林类型精细分类[J]. 遥感信息, 2019, 34(02): 104-112.

[5]郭航,张晓丽.基于遥感技术的植被分类研究现状与发展趋势[J]. 世界林业研究, 2007(03): 14-19.

[6]胡杰, 张莹, 谢仕义. 国产遥感影像分类技术应用研究进展综述[J]. 计算机工程与应用, 2021, 57(03): 1-13.

[7]颜伟, 周雯, 易利龙, 等. 森林类型遥感分类及变化监测研究进展[J]. 遥感技术与应用, 2019, 34(03): 445-454.

[8]Fassnacht F E, Latifi H, Stereńczak K, et al. Review of studies on tree species classification from remotely sensed data[J]. Remote Sensing of Environment, 2016, 186: 64-87.

[9]马玥. 基于多源遥感信息综合的湿地土地覆被分类研究[D]. 吉林大学, 2018.

[10]崔璐, 杜华强, 周国模, 等. 决策树结合混合像元分解的中国竹林遥感信息提取[J]. 遥感学报, 2019, 23(01): 166-176.

[11]Alam S M R, Hossain M S. A rule-based classification method for mapping saltmarsh land-cover in south-eastern Bangladesh from Landsat-8 OLI[J]. Canadian Journal of Remote Sensing, 2020(2): 1-25.

[12]于婉婉, 徐凯健, 赵萍, 等. Sentinel-2影像红边谱段对不同生长期区域优势树种识别的影响[J]. 地理与地理信息科学, 2021, 37(03): 42-49.

[13]Persson M, Lindberg E, Reese H. Tree species classification with multi-temporal Sentinel-2 data[J]. Remote Sensing, 2018, 10(11): 1794.

[14]Förster M, Kleinschmit B. Object-based classification of QuickBird data using ancillary information for the detection of forest types and NATURA 2000 habitats[J]. Springer Berlin Heidelberg, 2008.

[15]Fu B, Wang Y, Campbell A, et al. Comparison of object-based and pixel-based Random Forest algorithm for wetland vegetation mapping using high spatial resolution GF-1 and SAR data[J]. Ecological Indicators, 2017, 73(FEB.): 105-117.

[16]王凯. 基于SPOT5森林资源分类研究[D]. 浙江农林大学, 2014.

[17]李小梅, 张秋良, 李增元, 等. 基于对象的CHRIS遥感图像森林类型分类方法研究[J]. 内蒙古农业大学学报: 自然科学版, 2010(2): 6.

[18]Fassnacht F E, Neumann C, Forster M , et al. Comparison of feature reduction algorithms for classifying tree species with hyperspectral data on three central European test sites[J]. IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, 2014, 7(6): 2547-2561.

[19]Shang X, Chisholm L A. Classification of Australian native forest species using hyperspectral remote sensing and machine-learning classification algorithms[J]. IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, 2014, 7(6): 2481-2489.

[20]李军玲, 庞勇, 李增元, 等. 机载AISA Eagle Ⅱ高光谱数据在温带天然林树种分类中的应用[J]. 东北林业大学学报, 2019, 47(5): 5.

[21]Sharifi A, Amini J, Sumantyo J T S, et al. Speckle reduction of PolSAR images in forest regions using fast ICA algorithm[J]. Journal of the Indian Society of Remote Sensing, 2015, 43(2): 339-346.

[22]Koch H B. Exploring full-waveform LiDAR parameters for tree species classification[J]. International Journal of Applied Earth Observation and Geoinformation, 2011.

[23]Na X, Zang S, Liu L, et al. Wetland mapping in the Zhalong National Natural Reserve, China, using optical and radar imagery and topographical data[J]. Journal of Applied Remote Sensing, 2013, 7(1): 609-618.

[24]Sothe C, De Almeida C M, Schimalski M B, et al. A comparison of machine and deep-learning algorithms applied to multisource data for a subtropical forest area classification[J]. International Journal of Remote Sensing, 2020, 41(5): 1943-1969.

[25]鲁续坤. 基于机载LiDAR和高光谱数据的树种分类及三维显示[D]. 电子科技大学, 2018.

[26]吴艳双. 基于机载高光谱和LiDAR数据的树种分类[D]. 北京林业大学, 2019.

[27]李明泽, 付瑜, 于颖, 等. 基于多时相SAR数据和SPOT数据的盘古林场林分类型识别[J]. 植物研究, 2016, 36(04): 613-619+626.

[28]Laurin G V, Puletti N, Hawthorne W, et al. Discrimination of tropical forest types, dominant species, and mapping of functional guilds by hyperspectral and simulated multispectral Sentinel-2 data[J]. Remote Sensing of Environment, 2016, 176: 163-176.

[29]温一博, 范文义. 多时相遥感数据森林类型识别技术研究[J]. 森林工程, 2013, 29(2): 14-20.

[30]Nelson M. Evaluating Multitemporal Sentinel-2 data for Forest Mapping using Random Forest. 2017.

[31]李振, 胡慧萍, 杨敏华, 等. Landsat-8多时相遥感影像亚热带森林分类[J]. 测绘与空间地理信息, 2018, 41(09): 147-149.

[32]Immitzer M, Neuwirth M, Böck S, et al. Optimal input features for tree species classification in central Europe based on multi-temporal Sentinel-2 data[J]. Remote Sensing, 2019, 11(22): 2599-2622.

[33]Gill T K, Phinn S R, Armston J D, et al. Estimating tree-cover change in Australia: challenges of using the MODIS vegetation index product[J]. International Journal of Remote Sensing, 2009, 30(6): 1547-1565.

[34]刘峰, 张贵. 基于GIS和RS的广州市森林植被分类研究[J]. 湖南林业科技, 2004(01): 15-17.

[35]任冲, 鞠洪波, 张怀清, 等. 多源数据林地类型的精细分类方法[J]. 林业科学, 2016, 52(6): 54-65.

[36]Le W, Sousa W P, Peng G, et al. Comparison of IKONOS and QuickBird images for mapping mangrove species on the Caribbean coast of Panama[J]. Remote Sensing of Environment, 2004, 91(3-4): 432-440.

[37]陈工, 李琦, 张彦南, 等. 多源遥感信息提取桉树人工林[J]. 浙江林业科技, 2018, 38(02): 78-87.

[38]王娜, 李强子, 杜鑫, 等. 单变量特征选择的苏北地区主要农作物遥感识别[J]. 遥感学报, 2017, 21(04): 519-530.

[39]刘家福, 李林峰, 任春颖, 等. 基于特征优选的随机森林模型的黄河口滨海湿地信息提取研究[J]. 湿地科学, 2018, 16(02): 97-105.

[40]Cheng K, Wang J. Forest type classification based on integrated spectral-spatial-temporal features and random forest Algorithm-A case study in the Qinling Mountains[J]. Forests, 2019, 10(7): 559.

[41]周小成, 郑磊, 黄洪宇. 基于多特征优选的无人机可见光遥感林分类型分类[J]. 林业科学, 2021, 57(06): 24-36.

[42]Maxwell A E, Warner T A, Fang F. Implementation of machine-learning classification in remote sensing: An applied review[J]. International Journal of Remote Sensing, 2018, 39(9): 2784-2817.

[43]Jollineau M Y, Howarth P J. Mapping an inland wetland complex using hyperspectral imagery[J]. International Journal of Remote Sensing. 2008, 29(12): 3609-3631.

[44]Berberoglu S, Yilmaz K T, Özkan C. Mapping and monitoring of coastal wetlands of Cukurova Delta in the Eastern Mediterranean region[J]. Biodiversity & Conservation, 2004, 13(3): 615-633.

[45]Furuya D E G, Aguiar J A F, Estrabis N V, et al. A Machine Learning Approach for Mapping Forest Vegetation in Riparian Zones in an Atlantic Biome Environment Using Sentinel-2 Imagery[J]. Remote Sensing, 2020, 12(24): 4086.

[46]Tian X, Chen L, Zhang X. Classifying tree species in the plantations of southern China based on wavelet analysis and mathematical morphology[J]. Computers & Geosciences, 2021, 151: 104757.

[47]张乃静, 侯瑞霞, 纪平. 基于遥感影像和二类调查数据的林地类型分类方法对比研究——以广西凭祥市为例[J]. 林业资源管理, 2017(04): 89-96.

[48]郭瑞霞. 基于多源数据的落叶松人工林识别研究[D]. 西安科技大学, 2019.

[49]韩婷婷, 习晓环, 王成, 等. 基于决策树方法的云南省森林分类研究[J]. 遥感技术与应用, 2014, 29(05): 744-751.

[50]李若楠, 欧光龙, 代沁伶, 等. 基于GEE和Landsat时间序列数据的香格里拉森林类型分类研究[J]. 西南林业大学学报(自然科学), 2020, 40(05): 115-125.

[51]郗延彪. 基于Sentinel时序数据和深度学习算法的森林树种分类研究[D]. 中国科学院大学(中国科学院东北地理与农业生态研究所), 2020.

[52]Hinton G E, Salakhutdinov R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.

[53]Kattenborn T, Leitloff J, Schiefer F, et al. Review on Convolutional Neural Networks (CNN) in Vegetation Remote Sensing[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 173: 24-49.

[54]Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25.

[55]Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[J]. IEEE Computer Society, 2014.

[56]Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. Computer Science, 2014.

[57]He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C] //Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 770-778.

[58]Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39(4): 640-651.

[59]Schiefer F, Kattenborn T, Frick A, et al. Mapping forest tree species in high resolution UAV-based RGB-imagery by means of convolutional neural networks[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 170: 205-215.

[60]Pearse G D, Watt M S, Soewarto J, et al. Deep learning and phenology enhance large-scale tree species classification in aerial imagery during a biosecurity response[J]. Remote Sensing, 2021, 13(9): 1789.

[61]Hartling S, Sagan V, Sidike P, et al. Urban tree species classification using a WorldView-2/3 and LiDAR data fusion approach and deep learning[J]. Sensors, 2019, 19(6): 1284.

[62]Flood N, Watson F, Collett L. Using a U-net convolutional neural network to map woody vegetation extent from high resolution satellite imagery across Queensland, Australia[J]. International Journal of Applied Earth Observation and Geoinformation, 2019, 82: 101897.

[63]滕文秀, 温小荣, 王妮, 等. 基于深度迁移学习的无人机高分影像树种分类与制图[J]. 激光与光电子学进展, 2019.

[64]王雅慧, 陈尔学, 郭颖, 等. 高分辨率多光谱遥感影像森林类型分类深度U-net优化方法[J]. 林业科学研究, 2020, 33(01): 11-18.

[65]许慧敏, 齐华, 南轲, 等. 结合nDSM的高分辨率遥感影像深度学习分类方法[J]. 测绘通报, 2019, (08): 63-67.

[66]孙晓敏, 郑利娟, 吴军, 等. 基于U-net的“高分五号”卫星高光谱图像土地类型分类[J]. 航天返回与遥感, 2019, 40(06): 99-106.

[67]杨建宇, 周振旭, 杜贞容, 等. 基于SegNet语义模型的高分辨率遥感影像农村建设用地提取[J]. 农业工程学报, 2019, 35(05): 251-258.

[68]杨蜀秦, 宋志双, 尹瀚平, 等. 基于深度语义分割的无人机多光谱遥感作物分类方法[J]. 农业机械学报, 2021, 52(03): 185-192.

[69]杨明星, 徐天蜀, 牛晓花, 等. 基于Sentinel-1A雷达影像的思茅松林蓄积量估测[J].西部林业科学, 2019, 48(02): 52-58.

[70]向海燕, 罗红霞, 刘光鹏, 等. 基于Sentinel-1A极化SAR数据与面向对象方法的山区地表覆被分类[J]. 自然资源学报, 2017, 32(12): 2136-2148.

[71]张宇. SAR图像地形校正及应用研究[D]. 武汉大学, 2016.

[72]Gitelson A, Merzlyak M N. Spectral reflectance changes associated with autumn senescence of Aesculus hippocastanum L. and Acer platanoides L. leaves. Spectral features and relation to chlorophyll estimation[J]. Journal of Plant Physiology, 1994, 143(3): 286-292.

[73]Barnes E M, Clarke T R, Richards S E, et al. Coincident detection of crop water stress, nitrogen status and canopy density using ground based multispectral data[C]//Proceedings of the Fifth International Conference on Precision Agriculture, Bloomington, MN, USA. 2000, 1619.

[74]Gitelson A A, Keydan G P, Merzlyak M N. Three-band model for noninvasive estimation of chlorophyll, carotenoids, and anthocyanin contents in higher plant leaves[J]. Geophysical Research Letters, 2006, 33(11): 431-433.

[75]Merzlyak M N, Gitelson A A, Chivkunova O B, et al. Non-destructive optical detection of pigment changes during leaf senescence and fruit ripening[J]. Physiologia Plantarum, 1999, 106(1): 135-141.

[76]Frampton W J, Dash J, Watmough G, et al. Evaluating the capabilities of Sentinel-2 for quantitative estimation of biophysical variables in vegetation[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013, 82: 83-92.

[77]欧阳伦曦, 李新情, 惠凤鸣, 等. 哨兵卫星Sentinel-1A数据特性及应用潜力分析[J]. 极地研究, 2017, 29(2): 10.

[78]Haralick R M, Shanmugam K, Dinstein I. Textural features for image classification[J]. Studies in Media and Communication, 1973, (6): 610-621.

[79]Genuer, Poggi, JM, et al. Variable selection using random forests[J]. Pattern Recogn Lett, 2010, 31(14): 2225-2236.

[80]Stoppiglia H, Dreyfus G, Dubois R, et al. Ranking a random feature for variable and feature selection[J]. The Journal of Machine Learning Research, 2003, 3: 1399-1414.

[81]Diaz-Uriarte R, Andres S. Variable selection from random forests: application to gene expression data[J]. Quantitative Biology, 2005.

[82]朱文泉, 林文鹏. 遥感数字图像处理:原理与方法[M]. 北京: 高等教育出版社, 2015.

[83]盛庆红, 肖晖. 卫星遥感与摄影测量[M]. 北京: 科学出版社, 2015.

[84]罗小波, 赵春晖, 潘建平, 等. 遥感图像智能分类及其应用[M]. 北京: 电子工业出版社, 2011.

[85]Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation[J]. Springer International Publishing, 2015.

[86]Badrinarayanan V, Kendall A, Cipolla R. Segnet: A deep convolutional encoder-decoder architecture for image segmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(12): 2481-2495.

[87]Chen L C, Papandreou G, Kokkinos I, et al. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 40(4): 834-848.