A DCNN Geographic Object Extraction Method for National Geographic Condition Monitoring

In national geographic condition monitoring, there are such problems as fragmentation on distributions of ground objects, small plot area, lack of label data and high cost of sampling. On the basis of comprehensive analysis of the existing deep learning methods, by improving the model of LeNet5 and analyzing the perception field, this paper presents a DCNN method which can achieve pixel-level remote sensing classification by using scene classification methods for national geographic condition monitoring. The results indicate that the model generalizes well, and the Kappa coefficient reaches 0.967, which holds a great practical application value.