先验阈值优化卷积神经网络的作物覆盖度提取算法

Crop Covering Algorithm using Convolution Neural Network Based on Prior Threshold Optimization

农作物的生长观测是农业气象观测的一个重要部分,作物的覆盖度反映了环境对作物综合影响的结果,传统的先验阈值分割法受作物图像中存在的田间杂物、下雨或施肥后的土地以及光照阴影影响较大,会存在误分割的情况,针对这些问题,本文研究基于深度学习的作物与背景的自动分割问题,提出基于RGB和HSI关系阈值法优化的卷积神经网络(RGB-HSI-CNN)的作物图像分割提取覆盖度方法,解决了光照、遮挡、阴影等影响,取得了平均98.3%的模型准确率和97.53%的像素准确率,为后续作物生长状况监测以及农作物病虫草害的识别、监测等提供了有力支持。

The growth of crops is an important part of agricultural meteorological observations. Crop coverage reflects the re- sults of the environmental impact on the crop. The traditional threshold segmentation method may result in misclassification due to sundries, light, shadows, fertilization and rain in the crop image. In this paper, to solve these problems, propose a method to extract crop coverage and obtain image segmentation based on convolution neural network optimized by RGB and HSI relation threshold method （RGB-HIS-CNN） . The average model accuracy is 98.3% and the pixel accuracy is 97.53%. It provides strong support for the monitoring of growth status and the identification of crop diseases and insects.