Study on sample unbalance in landslide recognition algorithm based on depth learning
WANG Lixia, XI Wenfei, SHI Zhengtao, ZHAO Zilong, QIAN Tanghui, ZHAO Lei, MA Yijie
2024, 0(5):
12-18.
doi:10.13474/j.cnki.11-2246.2024.0503
Asbtract
(
)
PDF (6696KB)
(
)
References |
Related Articles |
Metrics
Landslides are a common geological disaster that can cause significant property losses and casualties to natural ecosystems and humans once they occur. How to quickly and accurately obtain landslide information is crucial to disaster prevention and mitigation. Traditional deep learning methods depend heavily on the quality of landslide samples, but the quality of existing samples is uneven, and the impact of landslide sample imbalance on the performance of deep learning models is rarely considered. Aiming at the problem of how to improve model accuracy by improving sample quality, this paper proposes a Faster R-CNN landslide target detection method based on multi-source unbalanced samples starting from sample quality. By conducting integrated training on a variety of imbalanced samples, the impact of different samples on the comprehensive performance of the model is studied. The results show that:①The accuracy rate of the model is 85.16%, F1 score of 0.69, precision of 56.96%, recall of 86.58%, and the missed detection rate is 0.33 under the imbalance of difficult samples. After strengthening the sample quality, the accuracy rate increases by 2.04%. The precision increased by 4.29%, the recall rate increased by 1.71%, and the missed detection rate decreased by 0.04. ②Under the imbalance of positive and negative samples, the accuracy rate of the model is 96.03%,F1 score of 0.78, precision of 64.50%, recall of 97.15%, and the missed detection rate is 0.09. After adding difficult samples to participate in the training, the accuracy rate drops by 8.45%. The rate dropped by 6.93%, the recall rate dropped by 7.25%, and the missed detection rate increased by 0.18. Difficult samples have a greater impact on the overall performance of the model. By improving the quality of these samples, the model detection accuracy can be improved. Therefore, the method proposed in this article provides a reference for solving the problem of landslide data sample imbalance in deep learning.