The Research On Weed Recognition In Rice Fields Using UAV Remote Sensing Imagery

Rice is the most important food crop of China,which is meaningful to ensure the national food security.During the process of rice production,weed is the key factor to enfluence the quality and quantity of rice production.With the development of large scale planting and the extension of direct seeding fields,the damage caused by weeds becomes more serious.According to statistics,the yield losses due to weeds is up to 40%.Thus,effective weed management is meaningful to ensure the quality and quantity of rice production.Among all the weed control means,chemical controls is widely used because of its easy operation and good effects.However,most farmers do not consider the spatial distribution of weeds,and apply a uniform spraying over the whole field.In order to ensure the chemical effects,some farmers even increase the chemical dose and spraying times.Overuse of herbicides has caused negative effects on environment as well as crop production.Accurate spraying emphasis applying enough does of herbicide on the weed infested areas,and do not uses chemicals in the weed free areas.In this way,the weed management can effectively reduce the use of berbicides while enhancing the chemical effects,which can well address the problem caused by overuse of herbicides.In the context of accurate weed management,an accurate prescription map can provide decision support information for the spraying applications.Based on the related literatures,we proposed to use the UAV remote sensing for data collection,and use object based image analysis(OBIA)and deep learning methods for data processing.The UAV remote sensing platform can efficiently cover the whold field and produce the orthophoto;the OBIA and deep learning methods can perform the per pixel classification,which can generate the weed cover maps and prescription maps.This work selected two rice fields in early growth stages(seedling and tillering stages)as experimental sites.A multi-rotor UAV was used for data collection.Image mosaicking was performed using the positions and features of images,which generated the orthophoto of the rice fields.In order to avoid the exhaustion of computational resources during analysis process,the orthophoto was divided into smaller regions of 1000 × 1000 pixels.Our dataset was divided into traning set,validation set,testing set.The mean intersection over union(MIU)and overall accuracy(OA)were adoped as the metrics for accuracy.The OBIA method was used for weed recognition.The mutliresolution segmentation and k-means method were used for image segmentation;the color and texture features were extracted as the feature vector;the BP neural network,support vector machine and random forest were used for classification.The hyperparameter optimization of segmentation models was performed using the grid search strategy,and the classifiers’ performance on the validation set was evaluated and compared.Experimental results showed that the MIU of multiresolution segmentation on testing set was 66.8%,and the execution time for an 1000 × 1000 image is 6463.1 ms;the MIU of k-means segmentation on testing set was 66.6%,and the execution time for an 1000 × 1000 image is 2343.5 ms.The deep learning method was used for weed recognition.The fully convolutional network(FCN)was applied for per-pixel classification of UAV images.Four CNN models(Alex Net,VGGNet,Goog Le Net,and Res Net)were used as pre-trained models and transferred to our training set.The skip architecture,fully connected conditional random fields(CRF),and patially connected CRF were used as improvement methods.The performance improvement brought by one single improvement method and hybrid improvement methods were evaluated and compared.Experimental results showed that for the pre-trained models,the VGGNet obtained the highest accuracy with an acceptable speed;for the the improvement methods,the skip architecture and partially connected CRF can effectively increase the accuracy,while the combination of different improvement methods can further promote the classification resuls.Experimental results showed that with skip architecture and partially connected CRF,the FCN achived 80.2% MIU on the testing set,and the inference time is 326.8 ms.The weed cover map for the whole rice field was generated by merging the classification results of OBIA and deep learning methods,respectively.Next,the weed cover map was divided into smaller grids using chessboard segmentation.By comparing the weed density of each grid with a given weed threshold,each grid was classified as treatment area or non treatment area,which generated the prescription map for the whole field.Experimental results showed that the deep learning outperformed the OBIA method in terms of accuracy and efficiency.For the two experimental sites,the deep learning method(FCN with hybrid improvement methods)can generate the prescription map in 20 minutes;with the weed threshold of 0.0%,the accuracy of prescription map generated by deep learning method is 83.6%,and the herbicide saving is 35.5%.The experimental results of this paper showed that the UAV remote sensing can effectively capture the variance of weeds and crops,which may obtained high classification accuracy;the deep learning method consistently outperformed the OBIA method in terms of accuracy and effieciency.The approach applied in this paper can efficiently generate an accurate prescription map,which may provide useful support information for the accurate spraying applications.