Research On Extraction Method Of Coffee Planting Information From Multi-source Remote Sensing Combined With Deep Learning And Obiect-oriented

Remote sensing technology is more and more applied in the field of agriculture,which provides a lot of information for Remote Sensing Extraction of planting areas of different crops.However,coffee as a plateau characteristic cash crop vigorously developed by Baoshan City,Yunnan Province,there are still relatively few studies on coffee planting information extraction by domestic and foreign scholars based on remote sensing data.Moreover,due to the characteristics of large area,large regional differences and low economic benefits per unit area of agricultural production activities,it is difficult for traditional ground survey to accurately obtain the distribution of crops to meet the actual needs.Therefore,real-time remote sensing monitoring of coffee planting areas and accurate acquisition of planting information and distribution of coffee forests are of great significance to promote the development of China’s coffee industry.This study taking Lujiang town of Baoshan City as the research area,selects GF-2,GF-6 and Sentinel-2A remote sensing images in 2020 as the basic data source,the key phenological features of coffee information identification are analyzed through the NDVI time series curve of Sentinel-2A data.The phenological features are then used to calculate and analyze the feature information differences of different ground objects under different data sources and to construct multi-source optimal classification parameters.And uses four classification methods: support vector machine,random forest classification,object-oriented classification and the combination of CNN and object-oriented to extract coffee in the study area,and compare the effects of different data sources and different classifiers on the extraction effect.The results are as follows:(1)Taking sentinel-2 image as the basic data source,the NDVI time series change curve is constructed by calculating the NDVI images of 7 phases in the coffee crop growth season in Lujiang town in 2020.After analysis,the characteristic difference between coffee field and other features is the most obvious in May,so it is determined that the best identification time of coffee is May.(2)Based on the phenological characteristics of coffee forest,combined with the field measured data,establish the interpretation marks of various ground features,and analyze the spectral characteristic curves of seven kinds of ground features such as forest,coffee,building,shrub,bareland,water and cropland.It is found that the recognition of various ground features can be better realized in the two bands of red light and near infrared.(3)Coffee was extracted based on single phase GF-2 data.Using the gf-2 image of the best extraction period,based on the reflectivity of the extracted forest,coffee and other land types,the classification dataset is constructed by combining with spectral,texture and terrain features,and using support vector machines,random forests,and object-oriented classification for coffee information extraction.The research shows that the overall classification accuracy obtained by by the three classification methods is more than 90%,and the kappa coefficient is greater than 0.85.Among them,OA and kappa of object-oriented classification method are 92.1% and 0.919 respectively,with the highest accuracy,complete classification results and the best effect.However,The OA of the relatively random forest classification method is the lowest,and the identification results are relatively fragmented,mainly because the spectral characteristics of coffee,forest and cropland are too similar to be well identified,and the misclassification phenomenon is obvious.In general,the three classification methods all have the phenomenon of misclassification and omission in the identification process to a certain extent,but the object-oriented classification method has the best extraction effect.(4)Based on the remote sensing images of GF-2,GF-6 and Sentinel-2 in 2020,compare and analyze the effects of different data sources and classification methods on coffee extraction.The results show that: at the method level,Sentinel-2 and GF-2 data are more suitable for the object-oriented classification method,while GF-6 data,SVM has the best effect.At the data source level,Sentinel-2A data is the best for identifying coffee forests with large patches;GF-2 data is the best for extracting coffee forests with small patches.Although the identification of coffee forest concentrated distribution areas by the three data sources is highly consistent with the actual situation,the identification effect of coffee forests distributed in low-altitude and sparse areas is not very complete and the effect is not ideal,which is also the difficulty of remote sensing identification of coffee forests.(5)Coffee information extraction based on the combination of deep learning and object-oriented.Using the object-oriented classification method to roughly extract the coffee ground from the images of the three data sources,and then based on the Tensorflow framework,the CNN network is used to build a training model to accurately classify the three data sources.The results show that the OA and Kappa coefficient of CNN classification are higher,reaching more than 90%,which is significantly higher than SVM,RF and object-oriented classification methods.From the perspective of data source,the extraction accuracy of GF-2 data is the highest.The OA 95.6% and Kappa coefficient is 0.936.At the same time,it can also be found that the classification accuracy of Sentinel-2 data is comparable to that of GF-2 data,indicating that Sentinel-2 data can also be used as an effective data source for coffee information extraction to a certain extent.On the other hand,through comparative experiments,the object-oriented classification method has shown a good classification ability,but in order to extract coffee crops more accurately,a convolutional neural network model(CNN)will be added to the object-oriented classification method extraction,the classification accuracy of the three data sources has been improved to a certain extent.The classification accuracy of gf-2 data and gf-6 data has been improved by 3.5% and2.9%,The classification accuracy of sentinel-2 data is improved by 0.3%.To sum up,making appropriate training data by combining data sources and object-oriented classification method is conducive to better complete coffee information extraction and have better classification accuracy.