Research On Extraction Of Potato's Spatial Distribution Based On BP Neural Network

The effective and rapid acquisition of agricultural information is an important guarantee for agricultural scientific decision-making.In addition to the commonly used statistical sampling survey methods,with the continuous development of spatial information technology,remote sensing data has been widely used in crop planting information extraction research;the use of remote sensing technology to quickly and accurately implement the monitoring of crop distribution,growth and other changes is the urgent needs of the agricultural production and management department.China has a vast area and a large population.The three major traditional food crops have limited space for productivity promotion.Food security issues are particularly important and prominent.Potato has shown a strong momentum of development with its advantages of wide adaptability and large room for improvement in production.It has become a new way to ensure national food security in the new period.In early 2015,the Ministry of Agriculture officially launched the potato staple grain project.At present,crops monitored by remote sensing in China are mainly concentrated on rice and other gramineous crops as well as legumes such as soybeans.There are few studies on potato for solanaceous crops.In recent years,neural network technology has become an important method for crop spatial information identification with its high classification accuracy and wide application range.Therefore,this paper chooses BP artificial neural network classification method,and combines the characteristics of the input features determined by hyperspectral feature analysis.Landsat-8 images are used to extract potato in Jijia Town and Xinglong Town,Changchun City,Jilin Province.The main methods and results are as follows:1.Constructing a variety of hyperspectral indices to detect the spectral characteristics of potato and other cropsPotatoes,corn,soybeans,and rice were selected as the study objects in the key growth period of potato,and the differences in characteristics of hyperspectral curves were studied.To better describe the spectral differences between potato and other crops,we set up a hyperspectral reflectance difference index,a hyperspectral first derivative difference index,a hyperspectral red edge amplitude difference index,a hyperspectral curvature difference index,and a hyperspectral vegetation index difference index.The results are as follows:The spectral curves of potato,corn,soybean,and rice have obvious differences.The reflectance values of potato and corn have the most significant differences in the blue band near 480 nm,with a disparity index value of 67.866%.The maximum differential index values with soybean and rice were 49.068% and 57.559%,all located at the green peak position near 550 nm.The spectral derivative curves of crops were transformed by the first derivative,and the spectral difference between potato and other crops was significantly amplified,and the amplification was most significant in the near-infrared region.The maximum values of the hyperspectral curvature indices of potato,corn,soybean,and rice were all located near the wavelength of 750 nm,and the differential index values were 78.365%,63.471%,and 80.882%.Among the commonly used vegetation indices,the ratio vegetation index and the enhanced vegetation index can significantly distinguish between potato,corn,soybean and rice.2.To conduct hyperspectral differences analysis of different crops and determine the input characteristics of BP neural network classificationBased on common remote sensors,both the vegetation index and the spectral band reflectance can be used as input eigenvalues to discriminate potato and other crops.We select the Landsat-8 image of July 4,2016,which is closest to the time of the ground spectral curve measurement time.Combining with the results of the hyperspectral difference analysis,we remove the red band with smaller difference and select the ratio vegetation index with the greatest difference in vegetation index.The input feature is determined as 23567 band + RVI for parameter optimization training of BP neural network.3.To determine the neural network classification structureLandsat-8 data was used to construct a land cover classification model based on BP neural network with ENVI platform,which was applied to potato and other crops in the study area.Taking the characteristics of the 23567+RVI band as input,the classification parameters are continuously adjusted to determine the optimal classification network structure.The final determined parameter values are: the activation function is a logarithmic function,the initial weight is 0.1,the learning rate is 0.1,the momentum factor is 0.5,the RMS is 0.3,the number of hidden layers is 1,and the number of trainings is 350,the minimum output activation threshold is 0.4.Potato distribution results and accuracy verificationThe results obtained by classifying the optimal parameters were tested with the accuracy of the confusion matrix.The total classification accuracy achieved is 95.8675% and the Kappa coefficient is 0.9395.The classification accuracy of the potato is 89.18%.The multi-temporal Landsat-8 image of the crop during the growth period was selected for comparison and verification.The dates are July 4,2016,August 5,2016,and September 22,2016.The final result is that the Landsat-8 image classification accuracy is the highest on July 4,2016.The reason for the analysis is that the time is closest to the measured spectrum from the ground.Therefore,the spectral characteristics are most similar to the differences.The precision comparison between the multi-phase GF1 images and Landsat-8 image in the crop growth period was performed.The results shows that the classification accuracy of Landsat-8 image is much higher than the multi-phase GF1 image classification accuracy.The reason for the analysis is that the spectral information contained in Landsat-8 images is richer than GF1.Even if the spatial resolution is not enough,spectral information is more important for potato identification.Compared with the ISODATA method,the minimum distance method and the maximum likelihood classification method,BP neural network has the highest classification accuracy.These indicate that the self-organizing learning adaptability of BP neural network has a great influence on the improvement of classification accuracy,and is an ideal way to obtain the spatial distribution of potato by remote sensing.