Extraction Of Larch Plantations Machine Learning Within High Spatial Resolution Images

Larch is one of the three important coniferous timber forest species in China,which is distributed in Northeast and Inner Mongolia.It plays an important role in wood production,water conservation and the formation and maintenance of forest ecosystem.In the future,it will also play an important role in achieving the goal of carbon neutralization and carbon peak in China.Using remote sensing technology to solve the problems of large-scale spatial distribution information of larch plantations in China is difficult to obtain,data collection is difficult,high cost,time-consuming and so on;rapid acquisition of spatial distribution of larch plantations has important practical significance for forest monitoring planning and ecosystem function evaluation,and achieving the goal of carbon peak and carbon neutralization.In view of the difficulties in obtaining large-scale spatial distribution information,data collection,high cost and time-consuming of larch plantations in China,this paper takes larch plantations in Heilongjiang Province as the research object,based on Google Earth Based on engine(GEE)platform,combined with sentinel-2 remote sensing image and forest survey data,larch extraction was carried out in Mengjiagang forest farm and Daxinganling area of Jiamusi City,Heilongjiang Province.Aiming at Mengjiagang forest farm,this paper discusses the application effect of machine learning algorithm(random forest,support vector machine)and deep learning in the extraction of larch plantation,analyzes the influence of multi temporal and multi feature synthesis and feature optimization on the classification accuracy of different algorithms,further determines the suitable algorithm for Larch Plantation recognition,and establishes the regional larch plantation extraction model.Aiming at Daxing’anling area,which is widely distributed in natural forest area,the extraction and identification of Larch were carried out,and the extraction model of Larch Forest in natural forest area was established.The following two problems are solved:① it is difficult to obtain high-quality remote sensing image data that meet the research needs in the process of obtaining spatial distribution of larch plantations based on single remote sensing data application platform;a large number of remote sensing image data preprocessing and machine learning classification algorithm are time-consuming and inefficient.Finally,based on the workflow and machine algorithm model of spatial distribution information extraction of larch plantation established in this paper,the spatial distribution information of Larch Plantation in Heilongjiang Province is extracted,and the application effect of remote sensing technology in large regional larch plantation extraction is further discussed,which provides technical support for efficient monitoring of dynamic changes of larch plantation.The main conclusions are as follows.(1)Research on feature space construction and feature optimization.The overall accuracy of multi-temporal image combination classification of random forest algorithm is 94.59%,and the Kappa coefficient is 0.92;the overall accuracy of multi-temporal image combination classification of support vector machine algorithm is 92.35%,and the Kappa coefficient is 0.91.Compared with single-phase image data,the multi-temporal features The combination can obtain more abundant spectral information of larch plantation,which makes the random forest and support vector machine algorithms have a greater accuracy improvement;in the combined classification of multiple feature factors,the image classification after texture features and terrain factors are added The accuracy has been improved and the classification results are more detailed;the image combination of the growth period and the defoliation period of larch can reflect the unique phenological law of larch.It is the best time combination for larch identification,which is conducive to the high-precision extraction of larch.(2)Extraction of Larch Plantation Based on machine learning algorithm.Random forest algorithm has the best comprehensive effect,and has achieved good classification results in small forest farm experiment and large regional model application,with the highest classification accuracy,the average verification accuracy is 94%,and the kappa coefficient is 0.919;support vector machine algorithm has much lower accuracy than random forest algorithm in single temporal image classification with less feature variables,but after the combination of multi temporal and multi feature The algorithm is sensitive to texture information,increasing texture information can significantly improve the classification accuracy;compared with only using spectral features,the combination of spectral features and texture features can effectively improve the classification accuracy(3.82%);improving the image resolution can increase the classification accuracy of deep learning algorithm for larch plantation,but the algorithm model training is limited It takes much more time than random forest and support vector machine algorithm,so it is not suitable for Larch Plantation extraction in large area.(3)Research on extraction of large-area larch plantations based on the establishment of models.The classification and extraction accuracy of larch plantations in Heilongjiang Province reached more than 72%,and the kappa coefficient was 0.62.The model established in the study has a good application effect in the extraction of large-area larch plantation information.At the same time,the classification accuracy of the model can be further improved under the condition that more training samples can be obtained.