Area Extraction And Number Statistics Of Tobacco Planting In Mountainous Areas Based On Deep Learning

Tobacco is a special cash crop,and planting area and number of plants are two important parameters for yield estimation.At present,the acquisition of planting area and number of plants in tobacco yield estimation is still done manually,which is time-consuming and labor-intensive,with low efficiency and great human error.For mountainous areas with complex terrain,tobacco plots are finely divided and scattered,making statistical work more difficult.Therefore,it is of great significance to find a fast and accurate tobacco extraction method,which can provide technical support for the government departments to carry out macro-control and policy making.Based on the tobacco leaf monitoring project of Henan Tobacco Company,this thesis selects Xiaojie Township,Luoning County,the mountainous area of western Henan Province as the research area.In July 2020,the satellite image of Gaofen-2,the aerial image of UAV and the measured data on the ground were obtained.Mainly based on the improved U-Net semantic segmentation algorithm,tobacco planting area is extracted,and the YOLOv3 target detection algorithm based on embedded attention mechanism is used to count the number of tobacco plants.The main work and results of this thesis are as follows:1.Based on the improved U-Net model and Gaofen-2 satellite,the tobacco planting area in the study area was extracted.The original U-Net model was improved by reducing the number of network layers,replacing the loss function and feature fusion,and the tobacco data set was used for training,and the tobacco planting area in the study area was predicted.The results show that the accuracy rate of the improved U-Net model is 85.53%,F1 score is 0.87,kappa coefficient is 0.85,the actual tobacco planting area in the study area is 107.09 ha,the prediction result of the improved U-Net model is 111.14 ha,and the area accuracy is 96.36%,which meets the actual work requirements.Compared with the original U-Net model,the accuracy of area extraction of tobacco is increased by 12 percentage points,and the accuracy of area extraction of tobacco is also increased by 13 to 14 percentage points compared with the traditional object-oriented classification method.It shows that the improved U-Net model can realize the high-precision extraction of tobacco planting area in mountain areas.2.Based on YOLOv3 model with embedded attention mechanism and UAV images,the number of tobacco plants was counted.By embedding different attention mechanisms,the improved YOLOv3 model was obtained,and then the self-made data set was used to train the model.Finally,tobacco samples were identified and counted.The results show that:(1)The accuracy rate,recall rate and F1 score of the original YOLOv3 model are 98.25%,89.89% and 0.94 respectively,and the recall rate of the model can reach 95.85%,96.74% and 96.13% after three commonly used attention mechanisms of SENet,CBAM and ECA are embedded respectively.The timing of tobacco number system pays more attention to the recall rate(precision rate),which shows that the method of embedding attention can meet the actual demand better.Compared with the common target detection algorithms SSD and Faster-RCNN,the recall rate of the improved YOLOv3 model is about 26 percentage points higher.(2)The test sets were resampled by 1.5 times,2 times,2.5 times and 3 times respectively,and the test sets with 4.93 cm,6.58 cm,8.22 cm and 9.86 cm resolutions were obtained.Using CBAM-YOLOv3 model to test,with the continuous decline of resolution,the recall rate is also decreasing.When the resolution is 8.22 cm,the recall rate is 70.16%,and when the resolution is 9.86,the recall rate drops directly to 51.75%,which indicates that the resolution at this time can’t meet the requirements of tobacco identification.(3)Using the best CBAM-YOLOv3 model to count the number of plants in the sample plot.The actual number of tobacco plants in plot 1 was 4087,and the model detection result was 3968,with an overall accuracy of 97.09%.2,036 plants were actually planted in plot 2,and the model detection result was 1,921 plants,with an overall accuracy of 94.35%.It shows that the improved CCBAM-YOLOv3 model can realize the accurate statistics of tobacco plants in mountainous areas.In this thesis,the deep learning method is used to make statistics on the tobacco planting area in the study area and the number of tobacco plants in some sample plots by using Gaofen-2 satellite images and UAV images,which provides high-precision parameters for tobacco yield estimation and an effective method for management departments to quickly obtain tobacco yield.There are 46 figures,12 tables and 86 references.