利用無人機與機器學習於邊坡巡檢與分析之研究

A Research of UAV and Machine Learning on Slope Sliding.

水保設施因為多座落於山區且數量及分布廣泛，因此造成許多設施交通可及性低，進而使得在水保設施在例行性巡檢或是於災後防治管理的困難度大大增加，因此增進管理設施之效率便十分重要。此研究使用無人機(UAV)進行航拍，以克服水保設施中無法拍攝到的範圍以及達到大量拍攝的目的。並基於現有之對水保設施的異常樣態分級與維護建議上，以python 所建置的深度學習演算法中進行自動化辨識與維護建議。在辨識完成後，利用HSB 色彩空間進行異常樣態的面積概算。在自動辨識中正規化方法成效達到0.966 及神經網路之成效為0.959。在異常樣態面積概算部分，在使用HSB 色彩空間下概算，並與現有之面積概算方法比較，其誤差約為7%，此誤差經專家評估其影響性低。

Water and soil conservation felicities are located in mountain area and distributed widely, which is lead to spend more time on traffic, and there have the difficulty of the routine detection or post-disaster response management will be higher. This is the importance to improve felicities management. This research will use UAV to surmount the area that people can’t reach to take picture and achieve the propose of numerous photographing. In the other hand, this research would build an automatically process based on deep learning algorism from Python, in order to give an automatic classification and suggestion. Then, area of abnormal situation in case will use HSB(Hue, Saturation, Brightness) models to roughly calculated. In automatic identification, got 0.966points by Regularization process. And got points 0.959 by Neural Network. After expert of this filed checked these two scores, the identify level from automatic model is same with expert’s. In the section of area of abnormal situation rough calculation, using HSB representation model could get the colour value on each item, and compared to existed area calculated method, the error rate is close to 7%. Because of felicities detection belongs to large area measurement, it means to the influence of calculated error is very low by expert’s estimate.