Method For Rice Identification Based On EVI Temporal Features From Deep Learning In Complex Landscape Regions

Timely and accurate information of rice spatial distribution plays a vital role in crop output estimation,food security and global climate change.Traditionally,crop information acquisition mainly takes a field survey,which has high cost and poor accuracy,and is not conducive to further promotion.The rapid development of remote sensing technology has opened up new way to extract rice information,and has matured now.In southern China,due to the influence of topography,climate conditions and other factors,rice cultivation has complex planting structure and a scattered distribution.Besides,a large amount of arable land has been abandoned.Frequent clouds and rains in the south is also a great challenge for collecting remote sensing data.Hence,there still faces some challenges in rice accurate identification via remote sensing technology.This study proposes a method for accurate extraction of rice information which is suitable for complex landscapes,using HJ-1A/B multi-spectral images with high temporal resolution.The specifics as follow:(1)Considering the over-saturation of Normalized Difference Vegetation Index(NDVI),the Enhanced Vegetation index(EVI)is used as an important index to monitor vegetation growth.And the growth characteristics of different land types would be analyzed by constructiong EVI time series.Among them,rice showed significant difference during different phenological stages.(2)With the aid of deep learning,an accurate extraction method of rice information in complex landscape areas is developed based on EVI temporal features.As a research hotspot in the field of deep learning,convolutional neural network(CNN)is able to extract deep abstract features with the advantages of multi-layer structure,local connection and weight sharing,which has been widely applied in remote sensing classification.In this study,the rice remote sensing recognition model was built based on deep temporal features of EVI,with CNN as the main framework.Additionaly,transer learning was used to extract EVI temporal features for fully training the model.(3)For the purpose of verifying the performance and reliablity of the model,support vector machine(SVM)is selected to make a comparison,and then evalute the precise qualitatively and quantitatively.The results show that the identification of CNN are more accurate for eight patches in the study area.Comparing three accuracy values,it is found that CNN is better than SVM on the whole,with the values are 93.60% and 91.05% separately.Moreover,the proposed model also shows high accuracy in rice extraction.The introduction of deep learning models can effectively improve the classification,and it shows high reliability in rice classification researches with mid-resolution remote sensing images.And more challenges and opportunities can be met in the fields of crop classification by combining the deep learning technology and extraction of vegetation index temporal features in the future research.