The article is devoted to the study of the use of neural networks to improve the selection of plant stands in precision agriculture technologies. The study takes into account the complex aspects of sample selection, such as the speed of image acquisition, the effectiveness of assessing the state of mineral nutrition and soil moisture, etc. The use of neural networks makes it possible to automate and increase the accuracy of selection, improving the quality of the analysis of plant stands, subject to compliance with soil sample evaluation technologies. The obtained results indicate the prospects of implementing this approach in modern agriculture. neural network, precision agriculture, plant samples, image recognition, training, shooting with a 0000-0001-7682-280X (A.4); 0000-0001-9797-3551 (A.2); 0000-0002-2120-1552 (A.3); 0000-0001-6433-3566 (A.4); 0000-0003-
UAV
Modern real systems and processes to one degree or another have development in time, therefore, they are stochastic. This means that the characteristics that describe their functioning are probabilistic and are random variables. The values of these quantities, as a rule, are in a certain interval, which sometimes has clearly defined boundaries, and more often - the boundaries are indefinite, vague. For example, such boundaries are inherent in parameters that are predictive in nature. Moreover, the more distant in time the forecasting horizon, the less its accurate, i.e. accuracy of estimates of the boundaries of possible values of such parameters. Therefore, in such conditions, the use of fuzzy intervals is preferable. Declaring model parameters in the form of a fuzzy interval is a convenient form for formalizing imprecise values. It is psychologically easy to give a fuzzy interval estimation, and the carrier of a fuzzy interval is guaranteed to contain the value of the parameter under consideration.
Recently, fuzzy modeling has become one of the most active and promising areas of applied research
in various fields [
2023 Copyright for this paper by its authors. membership function is proposed. This article is a development of these studies. It discloses a method for modeling random variables, the value intervals of which are given in a fuzzy linguistic form. In this case, both the Gaussian membership function and the beta distribution are used.
The conditions for agriculture are becoming more and more difficult, taking into account climate
change, the disproportionate increase in the cost of fertilizers and political risks. Precision agriculture,
also known as precision agriculture or precision agriculture, is an approach to agriculture that uses
modern technologies such as satellite imagery, sensors, geographic information systems (GIS) and
others to collect and analyze data about soil, plants and other factors of agriculture. The main goal of
precision agriculture is to optimize the use of resources (productive soil, water, fertilizers, etc.) and
maximize the yield while simultaneously reducing the negative impact on the environment.
Implementation of the concept of precision agriculture requires solving many organizational and
methodical issues and fundamentally improving the culture of production in the agricultural industry in
general and in crop production in particular. Thus, it is necessary to ensure the accumulation and
processing of large data sets, that is, to implement information technologies at a new level within the
limits of not a separate field, but an economy or even an industry, as shown in D. Yuniarto et al (2020)
in [
The implementation of Internet of Things technologies is promising, the experience of which is
shown in T. Wiangtong and P. Sirisuk (2018) in [
Sampling is advisable to be carried out using robots, since means of accurate positioning are necessary
in any case, since visual positioning according to landmarks in the field is not always possible at all
Winterhalter, W. et al (2020) in [
In general, considerable attention has been paid to the use of robots in agricultural production. Thus,
methods of route selection based on various optimality criteria have been developed in the works Y. R.
Milijas et al. (2021) in [
For the most common unmanned aerial vehicles with electric motors, which are easier to control due
to the absence of electromagnetic interference generated by internal combustion engines, the issue of
power supply is also well studied. These are the optimal routes in the conditions of limited storage
batteries, described in the work of N. Pasichnyk, et al (2021) in [
The issue of remote moisture assessment is extremely important for crop production, and the review
work by M. J. Pandian and D. Karthik (2022) in [
Research was conducted on production fields in 2019-2020 in Boryspil district of Kyiv region with coordinates 50º16' N, 30º58'E 50.0347. The Slantrane 3p system mounted on the base of the DJI Matrice 600 Pro UAV was used for spectral research. Data on separate spectral channels and vegetation indices calculated by the Slantview program were considered. The maximum detail (GSD 0.04 m/pixel) was obtained from the image window of the Slantview software (available variants of the NDVI index Green, Red and RedEdge). Monochrome images were used when studying the results of individual spectral channels (window of images), which were stored in bmp format to ensure completeness of information (Fig. 1).
To approximate the experimental data, the amplitude version of the modified Gaussian function (hereinafter GaussAmp) was adopted, without a shift along the ordinate axis. The choice of the Amplitude version of Gaussian peak function is due to the fact that, compared to the classical Gaussian function, it better describes the peak values and it is easier to adapt it to the variable size of the experimental area, which is important for the industrial implementation of solutions (1): , (1) where: N is the number of measurements (in our case, the number of pixels); X is the intensity of the color component, A is the amplitude; xc – average value; w is the standard deviation (corresponds to the long for production use. Accordingly, it is expedient to use the possibility of the Slantrange complex and its standard vegetation indices, since for them the calculations with the Slantview proprietary software for a field of 60 hectares took 40-50 minutes. The Slantview program interface provides access to standard vegetation indices, such as various variations of the most common NDVI index, as well as several proprietary indices such as Stress, Veg.fraction, etc., whose calculation formulas are not disclosed. The GreenNDVI and RedNDVI indices were selected for research. The results are presented in fig. 2. 200 150 s l e x i p f 100 o r e b m u N 50
0 200 150 s l e x i p f 100 o r e b m u N 50 0
GNDVI*1 w=0.007 GNDVI_w w=0.013 RNDVI*1 w=0.019
RNDVI_w w=0.019 0,3
0,4
0,5 GNDVI*1 w=0.007 GNDVI_w w=0.013 RNDVI*1 w=0.019 RNDVI_w w=0.019 0,3 0,4
0,5
Based on the results of the conducted research, it can be stated that the characteristics of the Gaussian distribution for the pixels of the NDVI vegetation index distribution map are significantly different from those obtained directly from the spectral channels. Thus, for NDVI indices, the standard deviation of the distribution in normal plants was equal to or even smaller than in those with a better moisture regime, in contrast to the results obtained directly based on the use of spectral channels. At the same time, the coefficient of determination for the distribution of NDVI indices was 0.85-0.95, which is significantly less than the distribution based on the results of using green and red spectral channels, where this indicator was 0.98 and higher. The red spectral channel and its derivatives in the form of indices turned out to be the most promising for identifying increased wet provision.
That is, as before the consideration of individual spectral channels or their combination in vegetation indices turned out to be insufficient for confident identification of correct samples for laboratory analysis.
A possible solution was proposed to use neural networks to analyze the distribution of areas in the field, since puddles are mostly circular in shape, which can be recognized in the field. In this case, there are no restrictions on the nomenclature of available indexes that can be used for analysis.
Convolutional Neural Networks (CNNs) are a powerful class of deep neural networks specially designed for processing grid-structured data such as images and videos. Convolutional neural networks are based on two main concepts: convolutions and pooling. Mathematically, convolution for 2D data can be defined as follows: ( , ) = ( ∗ )( , ) = ∑ ∑ ( − , − ) ∗ ( , ), (2) where I is the input data, K is the kernel (filter), S is the output data.
To achieve greater instability to various displacements in images, convolutional neural networks often use stride and padding operations. Print defines the step at which the filter moves over the input data, while padding adds extra pixels around the input data to help preserve dimensionality after convolution. Convolutional neural networks are often combined with fully connected layers to perform classification, regression, and other tasks. The results of the convolutional layers are concatenated and fed to the input of the fully connected layer. Convolutional neural networks have shown significant achievements in many areas where it is important to analyze large volumes of visual data. Their success is due not only to the power of the model, but also to the ability to learn abstract levels of representation of the hierarchical structure of input data, which makes them an indispensable tool for many tasks of analyzing and processing objects in large data sets, which allows considering this tool in the question of determining the optimal sampling point soil samples. Error backpropagation is based on gradient descent, where the loss gradient (a function that measures the difference between predicted and actual results) is calculated with respect to the network weights and parameters. This gradient shows how the weights and parameters should be changed to reduce the error. After calculating the gradient, the network applies an optimization algorithm (eg, stochastic gradient descent) to make corrective changes to the weights and parameters. This process is repeated for many data packages (mini-batches) during several training epochs. An important component of training convolutional neural networks is the use of a loss function, which measures the amount of error between predicted and actual results.
In summary, training convolutional neural networks involves passing data through the network, calculating the error, calculating the gradient, and adjusting the weights and parameters to minimize the error during training. Training of a convolutional neural network in Python for recognizing circles (puddles) in images was carried out on the fields obtained as a result of UAV surveying (Fig. 4). For this, the TensorFlow library was used, which allows you to easily build and train neural networks. Part of the network training code is shown in Fig. 5. A convolutional architecture with three convolutional layers and pooling followed by a fully connected layer for classification was used.
To train the model on the task of binary classification, the binary loss function binary_crossentropy was used. The testing of the network was carried out in order to identify areas with increased moisture supply, which were located in the fields and had a shape close to circles and differed in color. The testing of the network was carried out in order to identify areas with increased moisture supply, which were located in the fields and had a shape close to circles and differed in color. Accuracy Metrics: The first and most important metric is accuracy. The created network showed high accuracy on the test data: 0.80375 (Fig. 6), i.e., with such accuracy, the areas with high moisture content and cannot be used for obtaining soil samples were determined on the images of the fields. ROC curve and AUC: The receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) help determine the relationship between the sensitivity and specificity of a model. In an ideal case, the ROC curve will rise up to the left.
Network Testing
Fig. 7 shows the image of the curve, which allows us to establish the success of the network's training and its ability to recognize in the images areas that are not suitable for taking soil samples. Thus, the obtained distribution map was considered as an image, on which the objects defined as the remains of puddles were recognized using a neural network. Samples from these areas have different initial conditions and are limited in their suitability for mineral nutrition analysis.
5. Conclusions
1. The assessment of the nature of the distribution of both individual spectral channels and their combination in the form of vegetation indices turned out to be unprepared for the identification of uneven water supply of areas.
2. The red channel and its derivatives turned out to be the most promising in the direction of identifying the water supply of wheat.
3. The use of neural networks made it possible to identify probable areas with increased water supply on the maps of the distribution of vegetation indices in the field.
4. The duration of the identification using neural networks will not interfere with the sampling procedure, thanks to which such a procedure can be effectively implemented in agronomic practices.