Wildfire Detection Of Transmission Corridor Based On Himawari-8 Satellite

Yunnan Province is rich in forest resources,with complex terrain and meteorological conditions.Every winter and spring,with low rainfall and dry climate,a large number of wildfires occur.There are a large number of transmission lines in Yunnan,and wildfire seriously threaten the safe operation of transmission lines.At present,satellite wildfire detection algorithm of Yunnan Power Grid only uses a simple fixed threshold method,which has the problem of low accuracy.Further research on wildfire detection algorithms is needed to improve the accuracy of wildfire detection.Therefore,it has important research significance and practical value to carry out high-precision satellite wildfire detection of transmission lines.The Himawari-8 satellite data was used in this thesis to study the detection of wildfires in the Yunnan transmission corridor based on machine learning.The main research contents and results are as follows:(1)In view of the low accuracy of the current fixed threshold-based wildfire detection algorithm,the Ostu algorithm was used to calculate the optimal thresholds in this thesis,thus improving the fixed threshold cloud detection algorithm,and adding vegetation detection,solar flare detection,land cover type detection and other processing to reduce the false detection of fire points caused by related interference factors.Set the background window,adjust the window size according to the proportion of effective pixels in the window,calculate the average brightness temperature and standard deviation to identify the fire point,and improve the accuracy of wildfire detection.The experimental research results showed that the improved threshold method in this thesis has a precision of 0.57,an omission of 0.65,and an F1-score of 0.43.Compared with the traditional threshold method proposed by Xu,precision increased by 0.10,omission decreased by 0.02,and F1-score increased by 0.04,which reduces the missed detection of fire points to a certain extent and improves the accuracy of wildfire detection.(2)In view of the problem that the high time resolution of Himawari-8 data has not been effectively utilized,band7,band13,and band14 related to fire point detection were selected to carry out research on brightness temperature prediction based on historical time series data in this thesis.Five different deep learning time series prediction models were constructed,and the brightness temperature prediction research was carried out with single-band input and multi-band joint input.The experimental results showed that:1)in the research of single-band input brightness temperature prediction,the best brightness temperature prediction effect of band7 and band13 is the sequence to sequence(Seq2Seq)model,and its mean absolute error(MAE)is 0.30 K,0.26 K,mean square error(MSE)is0.22 K~2,0.34 K~2,and mean absolute percentage error(MAPE)is 0.10%,0.09%.The best brightness temperature prediction effect of band14 is the long short term memory(LSTM)model,and its MAE is 0.30 K,MSE is 0.42 K~2 and MAPE is 0.10%.2)the temporal convolutional network(TCN)model has the best brightness temperature effect in multi-band joint input prediction.MAE for brightness temperature prediction in band7,band13,and band14 is 0.29 K,0.26 K,and 0.28 K,MSE is 0.24 K~2,0.32 K~2,and 0.39 K~2,and MAPE is 0.10%,0.09%,and 0.09%.Compared with the results of single-band input,in the multi-band joint input brightness temperature prediction results,the MAE of band7decreased by 0.01 K,but the MSE increased by 0.02 K~2;the MAE of band13 remained unchanged,the MSE decreased by 0.02 K~2;the MAE of band14 decreased by 0.02 K,the MSE decreased by 0.03 K~2.Therefore,the overall prediction effect of multi-band joint input is better than single-band input.In the case of less cloud pollution,the TCN model can effectively use historical time series brightness temperature data to predict future brightness temperatures,and the predicted brightness temperature can be combined with the observed brightness temperature values at the same time point,as well as other spectral features,to participate in wildfire detection.(3)In view of the low detection accuracy caused by the relatively simple input features of machine learning for wildfire detection,four different strategies from the spectral,spatial,and temporal of Himawari-8 data were constructed in this thesis.The classic machine learning models were selected and combined with different strategies to conduct research on wildfire detection.The experimental results showed that:1)the strategy 4,which combines spectral,spatial and temporal information,and the random forest model have the best wildfire detection effect,with a precision of 0.62,an omission of 0.34,and an F1-score of 0.64.Compared with the threshold method,precision increased by 0.05,omission decreased by 0.31,and F1-score increased by 0.21,the average program running time of this method is 35.08 s.2)this method was applied to the detection of wildfires near the transmission lines of power grid,among the 295 real fire points,253 fire points were successfully detected,recall is 0.86.Among the fire points detected,90 fire points were detected at least 10 minutes in advance,and the earliest is 2hours ahead of time,the detection time of 163 fire points is consistent with the time found by Yunnan Power Grid,that proving the effectiveness of the wildfire detection method proposed in this thesis.