Crop Fine Classification Method Based on Grid Unit Feature Analysis

Precision agricultural technology is the general trend of agricultural modernization process, which needs to be based on efficient and accurate access to agricultural information. The category information and distribution information of crops provide important data support for agricultural monitoring and management. The classification and extraction of crops is the basic problem in agricultural research, and has important research significance and application value.

With the development of drones, UAV remote sensing technology has been widely used in agricultural remote sensing because of its low operating cost, high data accuracy and short acquisition period. However, it is difficult to improve the accuracy of image classification, because the UAV remote sensing data, which includes a large number of RGB three-channel images, is limited by insufficient spectral information. In particular, the spectral characteristics of various crops in the visible wavelength range are similar. It is difficult to accurately classify crops using remote sensing image analysis based on pixel spectral information.

Since the image of the drone has achieved very high spatial resolution, it is another effective method to use the image pattern recognition method to accurately identify and classify crops. The features of each type of crop are derived from image information, including spectral, texture, shape and context information.

Crop cultivation areas are characterized by human intervention. The similarity of various types of crops in the visible wavelength range is a typical confusing surface target extraction problem. Aiming at better solving the crop cultivation classification problem, this paper proposes a fine classification method for crops based on the analysis of grid cell features. The method uses the grid units acquired from the drone images to extract the features of continuous distribution. Then, the ResNet50, VGG16 and VGG19 models based on convolutional neural network (CNN) are used to classify each grid unit, achieving a high classification accuracy of 97.04%.