Research On Extraction Of Center Pivot Irrigation Crop Field In Africa Based On Deep Learning

Deep learning automatic analyzes and extracts data the feature in the process of the neural network training,and build a neural network that can simulate human’s thoughts.In view of the large amount of remote sensing image data,to apply deep learning into remote sensing image processing is an emerging problem.In agriculture field,cropland plays the most important role,which can provide important reference for agricultural yield forecast,resource allocation and agricultural economic schemes.Center pivot irrigation system（CPIS）uses a pipe on wheels pivoting around a central point to water crops via sprinklers.It often creates a circular crop field when viewed from above.This system requires expensive hydraulic pressure techniques and consumes huge energy,thus can make automatic irrigation come true and is widely applied in many countries and regions.Zambia in Africa is chosen as research area.Remote sensing data are multispectral image,gained from Sentinel-2A satellite during 2017 January to 2018 August with a lower cloud cover than 10%.Then these images are made a Pascal Voc data set,and use target detection networks to find center pivot irrigation cropland.And within the scope of the detection bounding box the croplands are extracted,which finally are made as tile maps,and published to the developed agricultural remote sensing information platform.The main works are as follows:1)Research traditional remote sensing image processing method that is used to extract the center pivot irrigation cropland,mainly including the variational region-based geometric active contour method,Hough transform.Two methods are combined to extract cropland,and problem is found that some croplands are extracted incorrectly.Besides a common problem is discussed that the methods of remote sensing images based on pixel is easily influenced by other pixels.2)Research the theories of neural network and target detection networks based on region proposal,including R-CNN,Fast R-CNN and Faster R-CNN.Then the source images are used to make a training dataset in format of Pascal Voc dataset,and Faster R-CNN is trained on this dataset.It was found that the miss detection for small cropland happened very often.3)Faster R-CNN is improved from the result of missed detection.In terms of the size of small cropland,the size of anchors are changed from the default{128^2,256²,512²}to{32²,64²,96²,128²},so that the small cropland can be detected precisely.After improved,the precision,recall,AP of test result respectively come to 88.13%,98.96%,90.66%。For the small cropland,it’s found that after improving Faster R-CNN,the recall of those croplands whose area is less than 1600px increase from 66.7%to 86.3%,and whose area is less than 2500px increase from 94.3%to 97.7%.4)The variational region-based geometric active contour method and Hough transform are used to extract cropland within the scope of detection bounding box,so as to shield the influence of other objects and correctly extract cropland.And the precision,recall of crop extraction result respectively arrive at 91.79%and 93.44%in 2017,and89.59%,94.02%in 20185)The extracted results are made into tile maps with GDAL,and then the agricultural remote sensing platform is developed with frameworks including Django,Openlayers,UI technology stack and so on.Finally the tile maps are published to this web platform.