Extraction Of Cultivation Area Of Mango From Remotelysensend Imageries Based On Deep Network Models

In modern agriculture,in order to meet the requirements of scientific management and meticulous operations,and to improve the planting efficiency of mango,there is a tremendous need to implement the comprehensive and accurate extraction of the cultivation area of mango.The increasing accessibility of high resolution remote sensing(HRRS)images makes it possible to accuratelyextract the cultivation area of mangoon a large scale.Due to the existence of excessive detail and texture information on HRRSimages,it is time-comsuming and laborious to extract the cultivation area of mango directly from from this type of images through manual interpretation,has high professional knowledge requirements for interpreters.For this reason,there is an urgent need to to develop efficient and automated methods to extract the the cultivation area of mango from remote sensed images.However,traditional methods are inadequate forextracting and interpreting visual features,and they often requirea large amount of prior knowledge to design features manually,thus resulting in the poor generalization ability.In order to address these problems,this thesis takes Renhedistricut in Panzhihua city as the research area,uses Google Earth HRRS images as the experiment data,and applies deep convolutional neural network models,which have achieved a remarkable success in the digital image analysis domain,to the extraction of the cultivation area of mango from remote sensing images.The proposed methods have been tested on HRRS images covering two different types of areas,i.e.,sparse vegetation area and dense vegetation area,confirming their effectiveness.To be specific,the main research contents of this thesis include the following aspects:(1)The extraction of the cultivation area of mango basedontheconvolutional neural network scene interpretation framework.Firslty,divide the Google Earth HRRS images to be analyzed in the experimental area into a series of overlapping image blocks,which constitue the basic units for the subsequent processing.Secondly,according to certain prior knowledge,a sample library correspondingto mango and non-mango regions based on the image block level is produced.Then,modelling on the convolutional neural network,the model parameters are estimanted based on the training samples,and followed by a forward propagation process to allocate scene category labels to various image blocks.Finally,the majority voting strategy based on the overlapping image block is used to ensure that each pixel in the original Google Earth HRRS image can be labeled with a unique class,thus achieving the pixel-level extraction of the cultivation area of mango.(2)The extraction of the cultivation area of mango basedonsemantic segmentation of the full convolutional neural network.According to certain prior knowledge,a sample library corresponding to mango and non-mango regions based on pixel-level annotation is produced.Then,under the framework of the semantic segmentation of the full convolutional neural network,the classical UNet model is trainined based on the training samples to obtain the corresponding model parameters.Finally,through a forward propagation process,the high-precision pixel-level extraction of the cultivation area of mango from the HRRS image to be interpreted is achived directly.The research achivements are of great significance forimproving the extraction precision of the cultivation area of mango from the HRRS images,and have potential application in the fine management of agricultural economic crops.