Distribution prediction of moisture content of dead fuel on the forest floor of Maoershan national forest,China using a LoRa wireless network

Bo Peng·Jiawei Zhang·Jian Xing·Jiuqing Liu

Abstract The moisture content of dead forest fuel is an important indicator of risk levels of forest fires and prediction of fire spread.Moisture distribution is important to determine wild fire rating.However,it is often difficult to predict moisture distribution because of a complex terrain,changeable environments and low cover of commercial communication signals inside the forest.This study proposes a moisture content prediction system composed of environmental data collected using a long range radio frequency band 433 MHz wireless sensor network and data processing for moisture prediction based on a BP (back-propagation)neural network.In the fall of 2019,twenty nodes for the collection of environmental data were placed in four forest stands of Maoershan National Forest for a month;7440 sets of data including temperature,humidity,wind speed and air pressure were obtained.Half the data were used as a training set,the other as a testing set for a BP neural network.The results show that the average absolute error between the predicted value and the real value of moisture content of fuels of Larix gmelini,Betula platyphylla,Juglans mandshurica,and Quercus mongolica stands was 0.94%,0.21%,0.86%,0.97%,respectively.The prediction accuracy was relatively high.The proposed distributed moisture content prediction method has the advantages of wide coverage and good realtime performance;at the same time,it is not limited by commercial signals and so it is especially suitable for forest fire prediction in remote mountainous areas.

Keywords Distributed moisture content prediction·Dead fuel·BP neural network

Introduction

As a basic component of forest ecosystems,dead fuel on the forest floor is an important material for forest fires and the main carrier of wildfire spread (Hu et al.2017),which directly affects the generation and spread of fires (Li et al.2016).The moisture content of dead fuel is the main determinant of fuel combustion properties and influences the rate at which the fuel reach their ignition point and the amount of heat released.It largely determines the intensity of fires and therefore it is an important indicator of fire risk (Wang and Luan 2014).

Wildfire scientists focus on the measurement or prediction of the moisture content of dead fuel on the forest floor.The most accurate method for the determination of the moisture content is drying and weighing.Sérgio et al.(2006)used this method to analyse seasonal moisture contents and fuels of different tree species in forests in central Portugal between 1996 and 2004.The relationship between moisture content and the occurrence of forest fires was investigated by studying the changes in moisture content of dead tree species after summer precipitation,thus establishing a safe timeline for the occurrence of forest fires.Oumar and Mutanga(2010) used the drying and weighing method to measure the moisture content of fresh leaves in the forest area of the Richmond area,Northern Cape Province,South Africa.Based on this,the feasibility of the South African satellite SumbandilaSat in assessing plant moisture was evaluated.Qi et al.(2014) used a partial least square regression method to measure the moisture content of combustibles in four forest locations near Missoula,Montana,U.S.A.The method was used as a means to obtain a calibrated moisture content to provide good modeling accuracy over time.The drying and-weighing method was used to determine the relationship between moisture content of fuel samples and plant physiological properties (Nolan et al.2018).They showed a clear link between physiological properties and seasonal variations in the moisture content of forest fuels.Maffei and Menenti (2019) developed and evaluated a new method to estimate the probability distribution function of a fire area in which fire duration and rate of spread varied with the index of vertical humidity.They noted a connection between the indexes and the moisture content of combustibles.In this study,the drying and weighing method was used to obtain the moisture content of combustibles to provide data support for the function model.Researchers in China have also made related studies.Wang et al.(2009) established regression equations for determining the moisture content of three fuel types using the method proving the relationship between fuel moisture content and climate,and established a fire risk prediction system for the Sehanba Dam.Zheng et al.(2013)used the drying and-weighing method to measure moisture content of forest fuels in Harbin Forestry Demonstration Base,and determined the fuel dynamics in the spring.Wang et al.(2020) measured fuel moisture content by this method and established a model of the relationship between moisture content and climate factors in Fengning County,Hebei Province.The drying and weighing method can accurately determine fuel moisture content but the method is time-consuming,difficult to measure in the field,and so it is as a calibration method for other methods.

The regression method of meteorological elements is a common method to predict the moisture content of dead fuel.It employs regression theory to establish a functional relationship between climatic data and moisture content of surface fuels.It is a relatively simple statistical model and is the main method for determining moisture content of surface fuels in China and is also the most widely used method globally (Zhang et al.2016).Marsden-Smedley and Catchpole(2001) reported on the results of a model for predicting the moisture content of surface fuels in Tasmanian grasslands.They developed a log-transformed model of moisture content versus relative humidity and dew point temperature through drying methods to continuously measure moisture content in spring,summer,and fall.Ferguson et al.(2002),using soil moisture sensors,measuredmoisture content over 180 days of fine dead fuel on the surface ofLarix gmeliniwoodlands in Florida,USA,giving a predictive model between the current day’s moisture content and the previous day’s,and 24-h rainfall.Alves et al.(2009) continuously measured the moisture content of fine dead fuels inPinus elliottiiEngelm.plantations in Brazil,and developed a prediction model between moisture content,air temperature,wind speed,and relative humidity.Slijepcevic et al.(2015)developed a predictive model for moisture content of fine surface fuels in eucalyptus stands in Tasmania,Australia.Two methods were used,one was the Canadian moisture code of fine dead fuels based on an empirical model and the other was on a process model.The results showed that both models resulted in good predictions in dry environments,but the process model outperformed the empirical one in wet environments.Schunk et al.(2017) conducted a study of moisture contents of fine surface fuels in eight different forest stands in southern Germany.Measurements were continuous over 10 h and the relationship between moisture content and various climate factors was fitted by Spelman’s correlation algorithm.This showed that moisture content prediction models varied considerably between coniferous and broadleaf forests due to different forest stands,and therefore the type of forest must be considered as an important factor when applying the models.Ruffault et al.(2018)reported on 15 years of observations on the moisture content of live fuels of six species of Mediterranean shrubs in 20 different locations in southern France.They evaluated six drought indicators,showing that a meteorological drought index used as a proxy for moisture content is a direct and extensive method.Zhao et al.(2021) conducted a modeling based on soil moisture and moisture content of fuels and compared the predictive value of the model with moisture content.This proved the process of the dynamic coupling of soil moisture andcoupling strength.Zhang et al.(2006)fitted multiple linear regressions to the relationship between moisture content of fine surface fuels and climatic factors in a cedar plantation at the Fujian Agriculture and Forestry University.Lu et al.(2010) studied the relationship between moisture content of fine surface fuels and climatic factors on Xishan Hill Beijing and obtained a prediction model of the moisture content of different forest stands.Ma et al.(2011)developed a regression analysis between the moisture content of yellow sumac leaves,pine needles,grasses,fine dead branches and soil,and climatic factors,showing that moisture content of the soil and temperature and humidity were more appropriate choices as predictors of moisture content of fuels.Jin and Zhou (2014) used climatic factor regression to show the relationship between moisture content of fine dead surface fuels and climatic factors in eight typical forest stands in Kunming,China and a prediction model was established.An in-depth study was carried out on the relationship between moisture content and climatic factors in Harbin Urban Forestry Demonstration Base,Harbin.The results showed that different climate factors had varying degrees of effect on the moisture content of different fuels(Zhou et al.2016).Zhang et al.(2020a) used the spatial heterogeneity of the moisture content of fuels to determine the speed of spread and direction of forest fires.The results showed that the spatial heterogeneity of fuel moisture content is best achieved by increasing the sampling density.In different fire prevention periods,different slope directions,sampling and location revealed that the research results can provide a scientific basis for forest fire prediction and prevention.Zhang et al.(2020b) used the relationship between fine fuel moisture content (FFMC) and climate to establish a prediction of the daily average fine fuel moisture content of northern China’s forests.They proposed a spline interpolation function to describe the daily changes in FFMC.After testing on-site measurement data,when the absolute error was <3% and <10%,the accuracy of the sunny slope model was 100% and 84%,respectively,while when the absolute error of the shade slope was <3% and <10%,the slope model was 72% and 76%,respectively.The results show that the sunny and shady slope models can predict daily variations of fuel moisture content.However,the meteorological factor regression method is inadequate and difficult to collect multi-point meteorological data,leading to uncertainty of the microenvironment of complex terrain.Using the traditional meteorological station to obtain climate data near the surface is difficult as the instrument is large,heavy,and difficult to disassemble,transport and install (Sun et al.2020).Data transmission is based on WiFi,ZigBee,4G and Bluetooth wireless technology,and the transmission distance is short,easily interfered with,expensive and cannot meet the demand of large-scale networking.The prediction range is limited by the specificity of the region,environmental factors and forest microclimate.It cannot be applied to a large scale range and it is difficult to promote the low universality application in other forest areas.

In order to solve these problems,this study proposes to develop a real-time prediction method for moisture content of forest fuels based on LoRa (long range radio),environmental factors collection devices and a BP neural network data processing system,to provide a new means to improve forest fire risk forecasting and wildfire management.

Material and methods

Overview of the experimental area

The experimental site is the Maoershan Experimental Forest Farm of the Northeast Forestry University,and the study period is the autumn of 2019.The forest farm is located near Harbin City (45°20’?45°25’N,127°30’?127°34’E).In the low mountainous and hilly area,the terrain gradually rises from south to north,with an average elevation of 300 m.The highest mountain is Maoer Mountain with an elevation of 805 m.The area has a temperate continental climate,with an annual average temperature of 2.8 °C,annual sunshine hours of 2471.3,a frost-free period of 120?140 days,annual precipitation of 723.8 mm,concentrated in June,July and August,annual evaporation of 1093.9 mm,and an annual average relative humidity of 70%.

The tree species includePopulus tomentosa Carr.,Phellodendron amurense Rupr.,Quercus mongolica,Juglans mandshurica,Betula platyphylla,andLarix gmelini.B.platyphylla,J.mandshurica,andQ.mongolicaare generally distributed in the middle and lower parts of the hillsides in various slope directions and degrees.Due to the relatively long growth period and high canopy density,canopy closure is proportional to the height of the tree,and the growth environment is wetting.Mixed tree species such asB.platyphyllaandP.tomentosasometimes form pure forests.Shrubs includeSyringa spp.andLonicera japonicaThunb.Understory herbs include grasses and ferns.This type of environment contains less flammable ground cover,where the litter is light in weight,loose and less compact,and is a medium flammable type.L.gmelinistands are pure plantations generally located in the middle and lower parts of the mountains,and the environment is relatively humid.There are a few shrubs and herbs under the canopy.The load of surface combustibles is the largest among all fuel types.TheL.gmelinineedles on the surface are small,in thick,tightly arranged humid layers,and are relatively inflammable.However,under extremely arid conditions,this ground cover can support a wild fire,and may lead a crown frie.The research fuel materials are the deciduous needles ofLarix gmelina,the leaves ofB.platyphylla,J.mandshuricaandQ.mongolica.

Practicability and feasibility of the e LoRa network for environmental data collection

Environmental data such as temperature of litter on the forest floor,humidity,atmospheric pressure and wind speed are the most important factors affecting moisture content.In recent years,LoRa has been rapidly adopted in areas such as smart meter reading due to its advantages of long transmission distance (more than 5 km),low power consumption and no commercial frequency band.Forest areas often have the problem of low coverage of commercial communication signals,so the advantages of LoRa are particularly suitable for forest areas.

Based on SX1278 LoRa wireless transmission,a network of 20 environmental collection nodes was established.Each node consisted of solar panels,power supply modules (3.7 V batteries),sensor modules (temperature,humidity,wind speed,pressure),GPS positioning module,STM32F103 ZET6 main control module,and 433 MHz 35dBi gain antenna.The diagram of the LoRa environmental factor acquisition device is shown in Fig.1.

Fig.1 Structure diagram of LoRa environmental factor acquisition device

The device can collect data of temperature,air humidity,atmospheric pressure,and wind speed with intervals of data collection continuously for 2 h by each collection terminal.The climatic data and data of information of geographic location are transmitted to the Central Processing Unit(CPU) and then to the background host computer through the LoRa data transmission module.When the data is sent,the LoRa module will automatically enter the sleep mode until activated during the next measurement,thereby achieving maximum energy saving.In order to verify the measurement accuracy of the instruments,the predicted results will be compared with actual data,and the measurement results of the meteorological collection device will be analyzed in the discussion section.The map is shown in Fig.2.

Fig.2 Experimental diagram of LoRa environmental factor acquisition device

BP neural network

The basic principle of a BP neural network is that the input signal acts on the output node through an intermediate node (implicit layer),which undergoes a nonlinear transformation to produce the output signal.The network prediction process is divided into network training and network testing.Each sample of network training includes input vector,expected output and the deviation between the network output value and the expected output value.By adjusting the value of the connection strength between the input node and the implicit layer node,and the connection strength between the implicit layer node and the output node as well as the threshold value,the error decreases along the gradient.Following repeated learning training,the network parameters (weights and thresholds)corresponding to the minimum error can be determined,after which the training stops.Network testing is a trained neural network that processes the input information of a similar sample in order to produce non-linear transformed information with minimal error.The process of predicting the moisture content of dead fuels on the forest floor involves taking temperature,humidity,wind speed,and barometric pressure as inputs and moisture content as an output.The nonlinear mapping between the two is determined through a BP neural network training process,and the data of the predicted moisture content is produced by the network training.

A BP neural network is a typical multilayer forward network composed of three parts:input layer,implicit layer,and output layer.In the prediction of moisture content of dead fuels,four types of parameters are selected as the input layer,and the number of neurons in the input layer is foremost.The moisture content is the output layer,so the number of neurons in the output layer is one.Input and output are the basic components of the network.In addition,the structure and intermediate function of the BP neural network are necessary conditions directly influencing the moisture content of fuels.An error analysis is made according to different network structures and functions combined with input and output data,and then the network structure and intermediate function that leads to the least error in the prediction results are selected.In the process of establishing the BP neural network,the network structure activation functions are screened according to the overall experimental data to ensure that the network prediction results reach the ideal level.The error calculation is based on the prediction results of the BP neural network,and the prediction accuracy of the network is evaluated to ensure the reliability of the prediction results.

Architecture of BP neural network

The number of nodes in the hidden layer affects the accuracy of the network and the efficiency of the training process.Too few nodes lead to learning difficulties,while too many lead to long training times and do not necessarily improve the accuracy.The number of hidden layer nodes is usually obtained through empirical formulas and multiple attempts.According to Kolmogorov’s theorem (Lin et al.2010),Yis set as the hidden layer andXis the number of neurons as shown in Eq.1.

After numerous attempts to obtain the average absolute error,the best training effect can be achieved when the number of nodes in the hidden layer is 10.The average error is shown in Fig.3.The mean absolute error (MAE) in the figure is the mean square error.In summary,the BP neural network model with a 4–10-10–1 structure is used to predict the moisture content of fuels (Fig.4).The MAE,which is the mean absolute value of all deviations of a single true value from a predicted value,can be used to measure the accuracy of prediction so as to precisely reflect the error of prediction because it can avoid the problem where the errors cancel each other out (Jia et al.2009).The average absolute error formula is shown in Eq.2.

Fig.3 Average absolute error results of different nodes in the hidden layer

Fig.4 Network model structure diagram

where,MAEis the average absolute error,n=20 environmental factor collection nodes,E iis the absolute error (Eq.4),R iis the true value,andR i^is the predicted value.Using experimental data to calculate the MAE of different network structures and functions,the most suitable network data is selected to predict the moisture content to obtain the most ideal results.

Activation function

There are three main types of activation functions commonly used in BP neural networks,namely,the linear activation function purelin,the logarithmic sigmoid activation function logsig,and the hyperbolic tangent sigma activation function tansig.According to the average absolute error of the three activation functions applied to the network (Fig.5),logsig was used as the activation function of the BP neural network.

Fig.5 The average absolute error results of different activation functions

Training function

Common training functions mainly include:steepest gradient descent function Traingd,momentum reversal gradient descent function tradingdm,adaptive gradient descent function trainingda,impulse gradient descent function traindx,and elastic gradient descent method trainrp.According to the five training functions applied in the average absolute error in the network (Fig.6),it was determined that the elastic gradient descent method trainrp would be used as the training function of the BP neural network.

Fig.6 The average absolute error results of different training functions

Experiment procedure

Twenty sets of terminal sensor equipment were arranged on the litter surface inL.gmelini,B.platyphylla,J.mandshurica,andQ.mongolicaforests,and 20 sets of on climate data generated in the four areas were collected every two hours.The climate data were transmitted for storage to the upper computer in the laboratory 1 km away via the LoRa wireless module of the terminal,which was synchronized with the terminal equipment of sensors.Every two hours withered leaves around the sensors were collected,weighed and recorded as wet weight.Climate data and litter were continuously collected for one month during the autumn frie prevention period.

The measurement of the sample data of the fuel moisture content was as follows:

1. Twenty sets of climate data in four areas were collected corresponding to the time when experimental materials(Dead fuel on forest surface) were collected;

2. The collected litter were brought back to the laboratory,put in a constant temperature drying box at 105 °C and dried to constant weight (according to the GB/T1931,a measure of wood moisture content in which the sample is dried at 103 ± 2 °C and considered to be fully dry when the difference between two weightings does not exceed 0.02 g).The dried litter material were weighed with an electronic balance and recorded as the dry weight;

3.The wet weight (M0) and dry weight (Me) were used to obtain the relative moisture content (M) as shown in Eq.3:

4. The climate data and moisture content of litter was imported into EXCEL;SPSS statistical software was used to filter the data and remove any data with obvious errors that do not conform to the overall change law.The remaining data was used as the to-be-trained sample set of the BP neural network (7440 groups of valid data remaining),and the overall experimental flow is shown in Fig.7.

Fig.7 Flow chart of the experiment

Results

Climate data collection

The terminal interface is shown in Fig.8.The interface of the host computer uses Labview to display visual meteorological elements.The transmitted data is read and received from the sensor terminal and the relay module using the LORA module connected to the USB interface and the VISA serial port.Climate data is called and then the climate data,the location of the collection devices and the prediction results are displayed on the interface of the host computer.When an abnormal situation occurs in a certain item of data collected from the forest,the program will give an alarm,shown in Fig.8 with devices 1,6,11,and 16 as a reference for demonstration.

Fig.8 Display interface of the host computer for four kinds of weather factors

Fig.9 Temperature data for twenty nodes collected every two hours over a month

Fig.10 Humidity data for twenty nodes collected every two hours over a month

Fig.11 Wind speed data for twenty nodes collected every two hours over a month

Fig.12 Atmospheric pressure data for twenty nodes collected every two hours over a month

Fig.13 Relative moisture content data for twenty nodes collected every two hours over a month

The results of the climate data and the relative fuel moisture content are filtered through SPSS data and imported into Matlab to obtain the results of temperature,humidity,wind speed,atmospheric pressure,and relative moisture content.Units 1 to 5 are arranged in theL.gmeliniforest,units 6 to 10 in theB.platyphyllaforest,11 ?15 units in theJ.mandshuricaforest,and 16?20 units in theQ.mongolicaforest.The expected devices 1,6,11,and 16 in individual stands are selected to display the climatic measurement results.The data results are shown in Figs.9,10,11,12,13.

From Figs.9– 13,it can be concluded that changes of temperature,humidity,wind speed and air pressure measured in the experimental area directly affected the moisture content of dead fuel on the forest floor.For example,on October 1st the weather was sunny,with high temperatures,low humidities,low wind speeds and low air pressure,so fuel moisture content decreased accordingly;in contrast,on October 4,the weather changed from showers to cloudy and the fuel moisture content increased due to low temperatures,high humidity,low wind speeds and high air pressure.Han et al.(2014),conducted a similar survey in Ningnan County,Sichuan Province and the results showed that variations in temperature,humidity,wind speed and other changes were inevitably related to surface moisture content.

BP neural network prediction results

According to the structure and parameters of the BP neural network described previously,there were 7440 sets of data,50% of which were used as training samples and the other 50% used as test samples.After 1000 training sessions,the results of the predicted moisture content were obtained.Among the 20 probe terminals,1,6,11,and 16 were selected to representL.gmelini,B.platyphylla,J.mandshurica,andQ.mongolica,respectively.The comparison between the predicted value of fuel moisture contents and the actual value is shown in Fig.14.

Fig.14 Comparison chart of predicted value and actual value of combustible moisture content

Discussion

By means of a self-made meteorological acquisition device to conduct real-time monitoring of temperature,humidity,wind speed,and air pressure,a neural network model was established to predict the moisture content of dead fuel on the forest floor ofLarix gmelina,Betula platyphylla,Juglans mandshurica,andQuercus mongolicastands.The significance lies in the establishment of a large-scale and high-density meteorological monitoring network to obtain real-time data of small environments in different locations.According to the changes of the data and the characteristics of the prediction model,a predicted value of moisture content with the minimum error was obtained by dynamically adjusting the weight of the prediction network.

By comparing the real and predicted values of the fuel moisture content (Fig.14),the mean absolute error could be obtained (Eq.2).This is the average absolute value of the absolute error between the real and the predicted values.This may be used to measure the accuracy of the prediction to accurately reflect the prediction error because it avoids the problem of errors cancelling each other out (Jia et al.2009).Among the 20 meteorological collecting devices,the devices 1,6,11 and 16,respectively,represent the absolute errors of the moisture content of the fuel in the four stands.The absolute error of the experiment is shown in Fig.15 found by Eq.4:

Fig.15 Absolute error result of moisture content between prediction and real values

where,Eis the absolute error,Ris the true value,andR^is the predicted value.

According to the absolute error results,the average absolute error of dead fuel in theL.gmelini,B.platyphylla,J.mandshurica,andQ.mongolicastands was 0.94%,0.21%,0.86%,and 0.97%,respectively (Hu et al.2018).Through research and comparison,the average absolute error of Nelson model is 9%.The results in this study are relatively reliable and the prediction accuracy more accurate than that of the random forest model because the experimental data were more fully prepared.On the other hand,the main purpose of this study is climate monitoring and dynamic model building,aiming to break through the limitations of traditional prediction models.

Conclusion

In this study,a distribution prediction system,based on LoRa wireless sensor and BP neural networks,was explored to realize the remote real-time and accurate prediction of moisture content of surface fuels for different forest stands.It provides richer and higher quality data of large area moisture content for enhancing forecasting forest fire risk and spread.