BP neural networks and random forest models to detect damage by Dendrolimus punctatus Walker

The construction of a pest detection algorithm is an important step to couple"ground-space"characteristics,which is also the basis for rapid and accurate monitoring and detection of pest damage.In four experimental areas in Sanming City,Jiangle County,Sha County and Yanping District in Fujian Province,sample data on pest damage in 182 sets of Dendrolimus punctatus were collected.The data were randomly divided into a training set and testing set,and five duplicate tests and one eliminating-indicator test were done.Based on the characterization analysis of the host for D.punctatus damage,seven characteristic indicators of ground and remote sensing including leaf area index,standard error of leaf area index(SEL)of pine forest,normalized difference vegetation index(NDVI),wetness from tasseled cap transformation(WET),green band(B2),red band(B3),near-infrared band(B4)of remote sensing image are obtained to construct BP neural networks and random forest models of pest levels.The detection results of these two algorithms were comprehensively compared from the aspects of detection precision,kappa coefficient,receiver operating characteristic curve,and a paired t test.The results showed that the seven indicators all were responsive to pest damage,and NDVI was relatively weak;the average pest damage detection precision of six tests by BP neural networks was 77.29%,the kappa coefficient was 0.6869 and after the RF algorithm,the respective values were 79.30%and 0.7151,showing that the latter is more optimized,but there was no significant difference(p>0.05);the detection precision,kappa coefficient and AUC of the RF algorithm was higher than the BP neural networks for three pest levels(no damage,moderate damage and severe damage).The detection precision and AUC of BP neural networks were a little higher for mild damage,but the difference was not significant(p>0.05)except for the kappa coefficient for the no damage level(p<0.05).An"over-fitting"phenomenon tends to occur in BP neural networks,while RF method is more robust,providing a detection effect that is better than the BP neural networks.Thus,the application of the random forest algorithm for pest damage and multilevel dispersed variables is thus feasible and suggests that attention to the proportionality of sample data from various categories is needed when collecting data.

Spatial variability of soil organic carbon and total nitrogen in the hilly red soil region of Southern China

To obtain accurate spatial distribution maps of soil organic carbon(SOC)and total nitrogen(TN)in the Hetian Town in Fujian Province,China,soil samples from three depths(0–20,20–40,and 40–60 cm)at 59 sampling sites were sampled by using traditional analysis and geostatistical approach.The SOC and TN ranged from 2.26 to 47.54 g kg-1,and from 0.28 to 2.71 g kg-1,respectively.The coefficient of variation for SOC and TN was moderate at 49.02–55.87%for all depths.According to the nuggetto-sill ratio values,a moderate spatial dependence of SOC content and a strong spatial dependence of TN content were found in different soil depths,demonstrating that SOC content was affected by both extrinsic and intrinsic factors while TN content was mainly influenced by intrinsic factors.Indices of cross-validation,such as mean error,mean standardized error,were close to zero,indicating that ordinary kriging interpolation is a reliable method to predict the spatial distribution of SOC and TN in different soil depths.Interpolation using ordinary kriging indicated the spatial pattern of SOC and TN were characterized by higher in the periphery and lower in the middle.To improve the accuracy of spatial interpolation for soil properties,it is necessary and important to incorporate a probabilistic and machine learning methods in the future study.

Decreasing the error in the measurement of the ecosystem effective leaf area index of a Pinus massoniana forest

Decreasing the forest ecosystem leaf-area index error(LAIe)helps accurately estimate the growth and light energy utilization of aboveground foliage.Analyzing light transmission in forest ecosystems can effectively determine LAIe.The LAI-2200 plant canopy analyzer(PCA)is used extensively for rapid field-effective LAI(LAIe)measurements and primarily to measure forest canopy LAIe values.However,sometimes this parameter must also be measured in forests with small clearings.In this study,we used the LAI-2200 PCA to obtain one A-value and four B-values each for the canopy,herbaceous layer,and forest ecosystem LAIe.Field measurements showed that the three LAIe types were obviously different.In certain quadrats,the average herbaceous layer(Dicranopteris dichotoma Bernh.)LAIe apparently exceeded that of the Pinus massoniana forest ecosystem.The sources of this error were measuring and recording A-value readings for small canopies and underestimating the ecosystem LAIe.We obtained similar coefficients of determination for both the pre-recomputation and post-recomputation of the canopy and forest ecosystem LAIe(R^2C 0.96 and R^2C 0.99,respectively);thus,the error was decreased.Measuring field LAIe with the LAI-2200 PCA and recomputation should compensate for LAIe underestimation in complex forest ecosystems.

Effect of layer thickness and voxel size inversion on leaf area density based on the voxel-based canopy profiling method

Voxel-based canopy profiling is commonly used to determine small-scale leaf area.Layer thickness and voxel size impact accuracy when using this method.Here,we determined the optimal combination of layer thickness and voxel size to estimate leaf area density accurately.Terrestrial LiDAR Stonex X300 was used to generate point cloud data for Masson pines(Pinus massoniana).The canopy layer was stratified into 0.10-1.00-m-thick layers,while voxel size was 0.01-0.10 m.The leaf area density of individual trees was estimated using leaf area indices for the upper,middle,and lower canopy and the overall canopy.The true leaf area index,obtained by layered harvesting,was used to verify the inversion results.Leaf area density was inverted by nine combinations of layer thickness and voxel size.The average relative accuracy and mean estimated accuracy of these combined inversion results exceeded 80%.When layer thickness was 1.00 m and voxel size 0.05 m,inversion was closest to the true value.The average relative accuracy was 92.58%,mean estimated accuracy 98.00%,and root mean square error 0.17.The combination of leaf area density and index was accurately retrieved.In conclusion,nondestructive voxel-based canopy profiling proved suitable for inverting the leaf area density of Masson pine in Hetian Town,Fujian Province.

A hybrid model for predicting spatial distribution of soil organic matter in a bamboo forest based on general regression neural network and interative algorithm

A general regression neural network model,combined with an interative algorithm(GRNNI)using sparsely distributed samples and auxiliary environmental variables was proposed to predict both spatial distribution and variability of soil organic matter(SOM)in a bamboo forest.The auxiliary environmental variables were:elevation,slope,mean annual temperature,mean annual precipitation,and normalized difference vegetation index.The prediction accuracy of this model was assessed via three accuracy indices,mean error(ME),mean absolute error(MAE),and root mean squared error(RMSE)for validation in sampling sites.Both the prediction accuracy and reliability of this model were compared to those of regression kriging(RK)and ordinary kriging(OK).The results show that the prediction accuracy of the GRNNI model was higher than that of both RK and OK.The three accuracy indices(ME,MAE,and RMSE)of the GRNNI model were lower than those of RK and OK.Relative improvements of RMSE of the GRNNI model compared with RK and OK were 13.6%and 17.5%,respectively.In addition,a more realistic spatial pattern of SOM was produced by the model because the GRNNI model was more suitable than multiple linear regression to capture the nonlinear relationship between SOM and the auxiliary environmental variables.Therefore,the GRNNI model can improve both prediction accuracy and reliability for determining spatial distribution and variability of SOM.

The co-effect of image resolution and crown size on deep learning for individual tree detection and delineation

Individual tree detection and delineation(ITDD)is an important subject in forestry and urban forestry.This study represents the first research to propose the concept of crown resolution to comprehensively evaluate the co-effect of image resolution and crown size on deep learning.Six images with different resolutions were derived from a DJI Unmanned Aerial Vehicle(UAV),and 1344 manually delineated Chinese fir(Cunninghamia lanceolata(Lamb)Hook)tree crowns were used for six training and validation mask region-based convolutional neural network(Mask R-CNN)models,while additional 476 delineated tree crowns were reserved for testing.The overall detection accuracy,the influence of different crown sizes,and crown resolutions were calculated to evaluate model performance accuracy with different image resolutions for ITDD.Results show that the highest accuracy was achieved when the crown resolution was between 800 and 12800 pixels/tree.The accuracy of ITDD was impacted by crown resolution,and it was unable to effectively identify Chinese fir when the crown resolution was less than 25 pixels/tree or higher than 12800 pixels/tree.The study highlights crown resolution as a critical factor affecting ITDD and suggests selecting the appropriate resolution based on the target detected crown size.

Developing a USLE cover and management factor(C)for forested regions of southern China

The Universal Soil Loss Equation model is often used to improve soil resource conservation by monitoring and forecasting soil erosion.This study tested a novel method to determine the cover and management factor(C)of this model by coupling the leaf area index(LAI)and soil basal respiration(SBR)to more accurately estimate a soil erosion map for a typical region with red soil in Hetian,Fujian Province,China.The spatial distribution of the LAI was obtained using the normalized difference vegetation index and was consistent with the LAI observed in the field(R^2=0.66).The spatial distribution of the SBR was obtained using the Carnegie-Ames-Stanford Approach model and verified by soil respiration field observations(R^2=0.51).Correlation analyses and regression models suggested that the LAI and SBR could reasonably reflect the structure of the forest canopy and understory vegetation,respectively.Finally,the C-factor was reconstructed using the proposed forest vegetation structure factor(Cs),which considers the effect of the forest canopy and shrub and litter layers on reducing rainfall erosion.The feasibility of this new method was thoroughly verified using runoff plots(R2=0.55).The results demonstrated that Cs may help local governments understand the vital role of the structure of the vegetation layer in limiting soil erosion and provide a more accurate large-scale quantification of the C-factor for soil erosion.