Abstract:
Background The supervision and management of soil erosion caused by construction projects is an important legal responsibility and social management function of the water administrative department. Satellite remote sensing imagery is an important method. As the basic data of supervision and management work, the disturbance patch boundary data of construction projects is currently mainly obtained by manual visual interpretation, which has low work efficiency and large cost investment.
Methods Based on the deep learning principle, a convolutional neural network was constructed to realize automatic identification of disturbance zone, so as to achieve the goal of automatic production of disturbance patch boundary data of construction projects. The remote sensing image of China's high-resolution No. 1 satellite with a resolution of 2 meters was used in this study. The constructed convolutional neural network contained 13 convolutional layers, 5 max-pooling layers, 1 global average pooling layer, 3 fully connected layers and 2 dropout layers. The activation functions in the middle and classification layer were RELU and Softmax, respectively. The training and testing samples were 5131 and 22923, and the proportion of positive samples in the training and testing sample sets were 15.38% and 4.69%, respectively.
Results The results showed that the optimizer, learning rate and batch size have a significant impact on the model training accuracy. The optimizer and learning rate have a negligible impact on the training time, and the batch size has a significant impact on it. In this study, the Adagrad optimizer, a 10-4 learning rate, and a 16 batch size were the optimal choices to obtain the best trained model with accuracy and loss of 0.9526 and 0.1670, respectively. The testing sample is used to validate the model, and the results showed, the average overall accuracy of the model is 97.52%, the average precision and recall of positive samples (classified samples of interest) are 72.44% and 83.90%, respectively, and the average F1 score is 77.73%. In general, the model recognition and classification results are basically consistent with the actual situation.
Conclusions This study provides a new method for the production of disturbance patch boundary data of construction projects, which greatly improves work efficiency, reduces input costs, and enhances the efficiency of soil and water conservation supervision and management. In the follow-up, the production of model training sample data should be strengthened, and the model parameters should be further revised to improve the accuracy of model application.
下载: