The comparative analysis of different two-wavelength laser altimeter versions for forest monitoring

This study presents a capability analysis of the various options for the technical implementation of a two-wavelength laser altimeter for monitoring forests. The results of woodlands pieces identification statistical simulation for a neuronic net using experimentally measured spectral reflectance are presented. It is shown that the neuronic net provides a strong probability of correct identification when using information about the reflectivity and the height of woodlands pieces. Two wavelengths 532 and 1064 nm and the neuronic provide the probability of correct identification of green broad-leaved and needle-leaved trees understory, wetlands and ground vegetation of greater than 0.89 and the misidentification probability of lower than 0.055.


Introduction
One of the actual tasks of monitoring the natural environment today is the monitoring of forests.Optical sounding is an effective remote monitoring method of the forest vegetation state [1][2][3][4].
Currently, in most cases, various vegetation indices are used for forest monitoring, based on measurements of multi-or hyperspectral sensing equipment in the visible and near IR spectral ranges.
Along with passive sensors, pulsed laser altimeters are used for forest monitoring (for example, [4][5][6]).Such laser altimeters generally operate at one wavelength.Data on the geometric characteristics of trees obtained by pulsed laser altimeters are combined with spectral information from passive sensors.
The advantage of using laser altimeters operating at several wavelengths is the ability to obtain qualitatively new information, for example, data on the woodlands ecologic system under the tree canopy [7].
In the simplest case, the laser altimeter at two wavelengths can be used.The paper presents the results of a capability analysis of the various two-wavelength laser altimeter versions for monitoring forests.

Problem description
The block diagram of a two-wavelength laser altimeter for monitoring forests is shown in Figure 1.
Here DM -dichroic mirror; M -mirror; IF -interference filter and FD -photodetector; OTS -optical transmission system; ORS -optical receiving system; ADC -analog-digital converter; IRPUinformation registration and processing unit.
The simplest versions of a laser radiation source for a two-wavelength pulsed lidar are a Nd: YAG laser (at the 1064 nm, 532 nm and 355 nm harmonics) and an optical parametric generator operating at 1570 nm (pumped by Nd: YAG laser).Such laser sources (see for example Ekspla NL200, NT242 [8]) have (at repetition frequencies from hundreds of Hz to units of kHz and pulse durations of 3-10 ns) pulse energy from units of mJ to hundreds of μJ, which allows airborne sounding at an altitude of several kilometers.The use of such laser sources makes it possible to implement a two-wavelength lidar at the wavelengths 355 and 1570 nm, 355 and 1064 nm, 355 and 532 nm, 532 and 1570 nm, 1064 and 1570 nm, 532 and 1064 nm.
Based on experimental spectral reflectance of various woodlands pieces, the paper conducts a statistical modeling in order to select the most effective wavelengths of pulsed laser altimeter for the task of monitoring forest areas.

Neuronic net for the analysis and processing the laser altimeter data
Mathematical simulation was carried out to evaluate the effectiveness of determining the woodlands pieces based on two-wavelength laser data.
A neuronic net was created to analyze the data of "measurements" of two-wavelength pulsed laser altimeter and classify the forest elements.The neuronic net architecture consists of three input neurons, one or two hidden layers (with intermediate neurons, the number of which varied from 5 to 15) and three output neurons.The neuronic net program was written in Python.The training of the created neuronic net was performed by backpropagation algorithm.The model optimizer -Adamax (modified Adam optimizer).
In statistical simulation, it was believed that the random noise of the laser altimeter receiving channel was distributed according to a Gaussian distribution with an average of zero and relative RMS of δ=1-10%.Statistical simulation was carried out on 10 3 realizations of random noise.The twowavelength "measurements" data was formed from the spectral reflectivity data of the woodlands pieces [9,10] by adding the random noise.
The probability of correct identification (correct determination of a woodlands pieces by its height and spectral reflection coefficient for selected sounding wavelengths) and incorrect identification (incorrect determination of a woodlands pieces) was assessed.

Statistical modeling results
The statistical simulation results of the probability of correct identification and the misidentification probability for the task of three different woodlands pieces separation (green broad-leaved and green needle-leaved trees understory, wetlands and ground vegetation) are given in Tables 1-7.
Tables 1-6 show the modeling results for pairs of wavelengths: 355 and 1570 nm, 355 and 1064 nm, 355 and 532 nm, 532 and 1570 nm, 1064 and 1570 nm and 532 and 1064 nm.
Two (out of three) input neurons received information about the spectral reflectivity of the forest area at two sounding wavelengths, while information about the height of the forest area (obtained from measuring the delay of the laser sounding pulse) was sent to the third input neuron.The neuronic net used for the modeling had five intermediate neurons in the hidden layer.The weights of these neurons were determined during the training process of the neuronic net using a training set.The three output neurons show the result of the neural network -the identification of pieces of forest: green broadleaved and needle-like undergrowth, wetlands, or ground vegetation.
During the training of the neuronic net, the reflectivity data of green broad-leaved trees, green needle-leaved trees, wetlands and ground vegetation [9,10] were divided into two halves: the training set and the validation set (for which the identification results are presented in tables 1-6 for random noise with a relative RMS of 5%).1000 training epochs were conducted at a training speed of 0.025.From the results presented in Tables 1-6, it can be seen that the best options, with the highest probabilities of correct identification and the lowest probabilities of misidentification of forest elements, are pairs of wavelengths 355 and 532 nm, 532 and 1570 nm, 532 and 1064 nm, 1064 and 1570 nm.However, in terms of simplicity of technical implementation, a pair of 532 and 1064 nm wavelengths has an advantage.
Two wavelengths 532 and 1064 nm along with the created neuronic net provide probabilities of correct identification of wetlands, understory of green needle-leaved and needle-leaved trees and ground vegetation above 0.89 and probabilities of misidentification below 0.055.
In addition to the Nd: YAG laser (with wavelengths of 1064 nm for the fundamental, 532 nm for the second, and 355 nm for the third harmonic) and an optical parametric generator at a wavelength of 1570 nm, a Ho: YAG laser at a wavelength of approximately 2.09 μm can be used for monitoring forest vegetation [8].
For a dual-wavelength laser, the use of 2.09 μm wavelength in the task of three different woodlands pieces identification: understory of green needle-leaved and needle-leaved trees, wetlands and ground vegetation, also leads to good results (see Table 7).However, the lidar version at 1064 and 2090 nm wavelengths is technically a more complex variant of a dual-wavelength lidar (than lidar at a pair of 532 and 1064 nm wavelengths).

Conclusion
Thus, a capability analysis of the various options for the technical implementation of a twowavelength laser altimeter for monitoring forests has been conducted.The results of woodlands pieces identification statistical simulation for a neuronic net using experimentally measured spectral reflectance are presented.It is shown that the neuronic net provides a strong probability of correct identification when using information about the reflectivity and the height of woodlands pieces.Two wavelengths 532 and 1064 nm and the neuronic net provide the correct identification probability of green needle-leaved and needle-leaved trees understory, wetlands and ground vegetation of greater than 0.89 and the misidentification probability of lower than 0.055.

Figure 1 .
Figure 1.Block diagram of a two-wavelength lidar.