Semantic Segmentation Algorithm Of Agricultural Remote Sensing Based On Deep Learning

Currently,with the rapid development of remote sensing,the timely and accurate estimation of crop classification based on remote sensing is of great significance for scientific and practical purposes.Crop yield estimation needs to be provided in a timely and accurate manner.It is a key issue to use multi-temporal remote sensing data effectively and flexibly.With the development of remote sensing and computer science,state-of-the-art research on crop classification has shifted from relying on a single static state to a combination of multiple spectral and time-series information.We propose a method to divide multispectral data into a single time phase in deep learning and then combine the information of the multi-temporal results on each neural network model.This can maintain the accuracy of the original information although the remote sensing multi-time series data is missing in some time phases.However,in actual practice,more accurate results can be further obtained with the superposition of new time-series data.We trained several semantic segmentation models for each individual time-series dataset.On the single time phase model,we obtained 85.9%-92.8% accuracy for classifying rice,77%-93% accuracy for classifying corn,and 77%-87.6% accuracy for classifying forest.Furthermore,our approach achieved 92.6% overall accuracy for classifying all catalogs by the multi-temporal verification dataset from May to September of 2017.The finding of our approach can be applied to the continuous iteration of agricultural yield estimation in agricultural insurance and crop growth change monitoring in big agricultural areas.