Precise Mapping And Spatiotemporal Analysis Of Paddy Rice Area In Complex Surface Landscapes

Farmland detection and variation are not only significant for food security and agricultural management,but also have crucial influence on agricultural production and economic development.Remote sensing,as the primary and burgeoning technology for data acquisition in earth observation,has been widely used in agricultural monitoring.In this study,a classification method combining phenological information with deep learning was proposed for precise rice mapping in the complex hilly region.Besides,we employed Landsat EVI time-series images to analyze the variation of paddy rice in Zhuzhou city.The major contributions can be concluded as follows:(1)Pre-training strategy is introduced to overcome the shortage of classical deep learning method,which reduces the amount of the training samples needed.The optimal network structure and initial weight are determined by analyzing the optimal parameters in the pre-training deep learning network.The results show that the pre-training method has better performance than the comparison methods in Zhuzhou,with overall accuracy of 82%.The pre-training strategy proposed in this study can effectively improve the efficiency of deep learning classification in small sample data set.(2)Phenological information are introduced into the pre-training deep learning network to precisely identify the rice field since single satellite image is difficult to distinguish the fallow rice field from abandoned land.Therefore,a hierarchical classification method combining phenological information with deep learning method was proposed.According to the test results,the proposed classification could effectively distinguish the abandoned land from the paddy rice field,with accuracy of 90%.(3)The spatial pattern evolution of rice field was explored according to spatial distribution clustering.For this purpose,the Mann-Kendall method was performed to analyze the breakpoint of Landsat EVI time-series images during 1987-2017,which is considered to be the number of land use changes.The landuse changes and possible driving factors are then analyzed.Local Moran’s I are employed to reflect the urban spatial pattern.(4)The results show three major changes over 31 years.From 1989 to 1991,the landuse changes greatly because a large number of people entered the city when five counties and four districts are established in Zhuzhou city.From 2002 to 2004,the landuse varied because of the human activity and the heavy industry development.From 2008 to 2012,the area of rice field decreased while that of abandoned land increased since the agricultural population in Zhuzhou has decreased with the economic development and the influence of “cadmium rice incident”.Additionally,the area of agricultural field peaked in 2000.The percentage of agricultural field increased from 1990 to 2000 before decreasing from 2000 to 2010.