当前位置: X-MOL 学术Environ. Earth Sci. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Modeling gully erosion susceptibility in Phuentsholing, Bhutan using deep learning and basic machine learning algorithms
Environmental Earth Sciences ( IF 2.8 ) Pub Date : 2021-04-02 , DOI: 10.1007/s12665-021-09599-2
Sunil Saha , Raju Sarkar , Gautam Thapa , Jagabandhu Roy

The present study attempts to demarcate the areas susceptible to gully erosion in Phuentsholing, Bhutan, using Deep Learning CNN (convolution neural network) and artificial neuron network (ANN), Support Vector Machine (SVM) and maximum entropy, three basic machine learning techniques in the GIS setting. Application of deep learning technique is new in the field of gully erosion. Considering the 240 gully pixels and seventeen gully erosion conditioning factors (GECFs), the gully erosion susceptibility maps (GESMs) were prepared. Out of the 240 gully pixels, 70% were used as training datasets and 30% were used as validation datasets for modeling and judging the GESMs. The GECFs were selected based on the previous literatures and multi-collinearity test. The importance of the GECFs was assessed by the chi-square attribute evaluation (CSEA) and random forest (RF) methods. Finally, applying the receiver operating characteristics’ area under curve (AUC-ROC), RMSE, MAE and R-index, the robustness of the GESMs was evaluated and compared. The GESMs were classified using natural break classification method into very high, high, moderate, low and very low susceptible classes. Nearly, 20% of the study area has very high susceptibility to gully erosion. As per the results of CSEA and RF methods, sand concentration, land use\cover and altitudes have the largest contribution in making the area very susceptible to gully erosion. Results of the validation techniques recognized the entire selected model as accurate and robust. Among the selected models, the capability of CNN model (AUC = 0.910, MAE = 0.029, RMSE = 0.171 for training data and AUC = 0.929, MAE = 0.089, RMSE = 0.299 for testing data) in predicting the gully erosion susceptibility is higher than other models. The produced GESMs will be helpful to the researchers as well as decision makers in establishing gully erosion management strategies.



中文翻译:

使用深度学习和基本机器学习算法对不丹丰特林地区的沟壑侵蚀敏感性进行建模

本研究尝试使用深度学习CNN(卷积神经网络)和人工神经元网络(ANN),支持向量机(SVM)和最大熵,不丹的三种基本机器学习技术来划分不丹Phuentsholing易受沟壑侵蚀的区域。 GIS设置。深度学习技术的应用在沟壑侵蚀领域是新的。考虑240个沟壑像素和17个沟壑侵蚀条件因子(GECF),绘制了沟壑侵蚀敏感性图(GESMs)。在240个沟壑像素中,有70%被用作训练数据集,而30%被用作验证数据集,用于对GESM进行建模和判断。根据以前的文献和多重共线性测试选择了GECF。通过卡方属性评估(CSEA)和随机森林(RF)方法评估了GECF的重要性。最后,应用接收机工作特性的曲线下面积(AUC-ROC),RMSE,MAE和R指数,评估并比较了GESM的鲁棒性。使用自然断裂分类法将GESM分类为极高,高,中,低和极低易感性类别。几乎有20%的研究区域对沟壑侵蚀非常敏感。根据CSEA和RF方法的结果,沙尘浓度,土地利用\覆盖度和海拔高度对使该地区极易受到沟壑侵蚀的影响最大。验证技术的结果将整个所选模型视为准确而可靠的模型。在所选模型中,CNN模型的功能(AUC = 0.910,MAE = 0.029,训练数据的RMSE = 0.171,AUC = 0.929,MAE = 0.089,RMSE = 0.299(测试数据)在预测沟壑侵蚀敏感性方面高于其他模型。产生的GESM将对研究人员和决策者建立沟壑侵蚀管理策略有所帮助。

更新日期:2021-04-02
down
wechat
bug