基于遥感和机器学习的内陆水体水深反演技术

文章主要根据机器学习算法（随机森林算法和极端梯度提升算法）和遥感水深反演的原理,利用Sentinel_2多光谱卫星数据和无人船实测水深数据,对内陆水体——梅州水库建立了随机森林（RF）、极端梯度提升（XGBoost）和支持向量机（SVM）水深反演模型,并对反演结果进行对比分析。结果表明：1）RF的训练精度为97%,测试精度为0.80;XGBoost模型的训练精度为97%,测试精度为0.79;SVM的训练精度为90%,测试精度为0.78。说明了在水深预测方面RF模型和XGBoost模型比SVM模型表现更好,对各个区段的水深值较为敏感。2）根据运行时间考察各个模型的效率,其中RF模型从读取数据至输出结果耗时3.92 s;XGBoost模型4.26 s;SVM模型6.66 s。因此,在反演精度和效率上RF模型优于XGBoost模型优于SVM模型,且RF模型的预测结果图细节更加丰富,轮廓更加分明;XGBoost模型次之,但总体效果也较好;SVM模型表现最差。由此可知,机器学习水深反演模型获得的水深结果精度明显提高,解决了传统水深反演模型精度不高的问题。

Estimating InlandWater Depth Based on Remote Sensing and Machine Learning Technique

This study investigates the application of three machine learning algorithms,e.g.the Random Forest(RF),Extreme Gradient Boosting(XGBoost),and Support Vector Machine(SVM)algorithms,in water depth estimation using Sentinel-2 multispectral satellite images.A case study is conducted in the Meizhou Reservoir in Guangdong,China,and a comparative analysis of the inversion results is performed.The results are validated against in-situ measured data using an unmanned ship,in which a global positioning system and a single wave velocity sounding system are integrated.The experimental results based on our water depth inversion models demonstrate good inversion accuracy and efficiency of the machine learning water depth inversion model constructed using 7925 water depth data samples and satellite multispectral images.By adjusting the key parameters of each model such that the model reaches the optimal state,the determination coefficient(R2),Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and Bias were used to evaluate the accuracy of the models.For water depth inversion,the RF model indicated that R^2=0.80,RMSE=2.21,MAE=1.51,and Bias=0.00;the XGBoost model indicated that R2=0.79,RMSE=2.22,MAE=1.50,and Bias=-0.01;and the SVM model indicated that R2=0.78,RMSE=2.23,MAE=1.60,and Bias=0.01.To determine the efficiency of each model,the models'running times were obtained:3.92 s for RF,4.26 s for XGBoost,and 6.66 s for SVM.Notably,the RF model is superior to the XGBoost and SVM models in terms of inversion accuracy and efficiency.Moreover,the advantages and disadvantages of each model can be inferred from the prediction result graph.The prediction results of the RF model are more detailed,and the terrain is clearer compared with those of the XGBoost model.The SVM model does not achieve ideal prediction using the same dataset,and the error at a shallow water depth is much worse than those of the other two models.The scatter plots indicate that the inversion deviation is not obvious between the RF and XGBoost models,and the prediction results are more consistent with the actual results.However,the scatter plot of the SVM has a horizontal bar near 11 m;therefore,it predicts the water depth value in this interval incorrectly.Furthermore,the errors of the RF and XGBoost models evaluated by the standard evaluation method are lower than that of the SVM model,indicating that those models have a higher prediction accuracy for water depth inversion.In summary,the machine learning models for water depth estimation yielded good performances in this study,and they are suitable for estimating the water depth using remote sensing images with lower economy and time cost,especially in unreachable waters.