Identification Of Rice Planting Areas Based On Enhanced Vegetation Index And Convolution Neural Network

Accurate and timely understanding of rice growing areas is essential for rice yield estimation,global climate change research and agricultural resource management.However,the traditional manual statistics method takes time and effort,which directly restricts the formulation of relevant agricultural policies.In recent years,remote sensing technology has been widely used in the field of agriculture.Because of its high resolution,wide observation range,strong timeliness and many other advantages,it has become an important technical means for agricultural monitoring,precision agriculture and other work.However,due to the large volume of data,obvious interference in similar areas,and variable target scales in high-resolution remote sensing images,the accuracy of rice region extraction is lower.When extracting rice planting areas based on medium-resolution image data,it is often limited by such factors as low image clarity and poor timeliness,and it is difficult to obtain accurate training sample datasets.To solve the above problems,this paper chooses Landsat-8 and MODIS medium and low resolution image data as data sources,chooses ENVI and Arc GIS as the main data processing platform,combines Enhanced Vegetation Index(EVI),extracts training samples for in-depth learning,and fine-tunes the pre-training convolution neural network(CNN)model based on the idea of migration learning through multiple iterations.To evaluate the accuracy of the model,a traditional machine learning method,Support Vector Machine(SVM),is introduced to compare the classification results.The purpose of this study is to explore the feasibility and practicability of in-depth learning technology in extracting medium-resolution remote sensing images of rice planting areas in a wide range of confusing complex landscape areas.The main conclusions are as follows:1)EVI is a common index for extracting types of ground objects in dense vegetation areas.It has a good distinction for obtaining spectral characteristics of the four types of ground objects.Using EVI to set specific thresholds to extract deep learning training samples is of great significance for reducing manual calibration and improving recognition accuracy.2)By using the method of migration learning,the CNN model built in this paper can effectively reduce the demand for training sample data.At the same time,by comparing the parameter optimization experiments,it is found that the two parameters of the network,batch data and learning rate,are set to 100 and 0.1 respectively,which makes the model more convergent and the fitting accuracy reaches above 0.90.3)By comparing the classification results,it is found that the overall classification accuracy and kappa coefficient of CNN are 88.79% and 84.92% respectively,and the overall classification accuracy and kappa coefficient of SVM are 86.36% and 81.67%respectively,which shows that CNN is superior to SVM in the overall classification performance.At the same time,by comparing the recognition ability of the two models on rice,it is found that CNN is also superior to SVM in both user accuracy and producer accuracy,The accuracy difference is up to 4%,which shows that CNN has better recognition ability for rice.4)Based on the CNN and SVM classification models built in this paper,the rice planting area in Chengdu in 2017 was identified as 128719 hectares and 150 587 hectares,respectively.By comparing with the statistical data,the accuracy of the two classifications was 94.30% and 89.68%,indicating that the CNN fits the statistical data better.