Internet of Things-Based Crop Forecasting System

Abstract

In this paper covered crop forecasting system. Even though there are a lot of precision farming software products available, many of the current innovations are not designed to guarantee the Russian Federation's food security or to provide data for additional analysis by experts who are not affiliated with the developer company. A system for the collection and subsequent analysis of aerial photography data has been built by us. In addition to providing methods for modelling the vegetation process using high-quality determinate mathematical models and forecasting a field's yield in the upcoming season, the paper also explains how data is presented, how vegetation indices are calculated for a given date, and how much this crop will cost. I discuss the possibilities of my system and show how to improve it further.

Specification

Description:Scientists in Russia and other nations are pointing out that precision farming requires the application of artificial intelligence techniques. However, the methods described in individual articles employ conventional procedures, and the software used for it these days frequently remains in the development stage and is not made available for purchase. The available literature on the application of AI techniques in precision farming is analyzed in the study. This analysis shows that Russia is far less advanced than numerous nations, including China, Australia, Canada, and the EU. The consequence of the efforts of research teams at Clark University (USA) is the creation of the Idrisi platform. About ten software modules have been created, and Idrisi GIS is one of them. It provides over 300 analytical tools for working with geographical datasets for globe exploration, with a primary concentration on raster data. A variety of tools for statistical analysis, change analysis, and surface analysis are among the unique aspects of the GIS Analysis Toolkit, along with a set of multi-criteria decision methods. Additionally, there are specialized graphical simulation platforms available for decision support and dynamic simulation. Additionally, the GIS Analysis toolkit offers a scripting environment and a very adaptable application programming interface (API) that enables the integration of Python, Delphi, and C++ modules. However, this program is not open source. On a different section of the website, the creators post open tasks. Scientific articles covering the work of this complex of programs could not be found. The needs of agriculture are at a considerable distance from the sole work dedicated to IDRISI. The One Soil platform, which bills itself as a free precision farming tool, is one of the most quickly evolving innovations at the moment. Currently a Swiss firm with offices in multiple EU nations, the project started in Belarus. An interactive map called One Soil Map shows information on 27 crops and 60 million farms across 43 European and American countries. A neural network is used to gather the statistics, while Mapbox was utilized for visualization. A farmer can use One Soil Map to see the number and size of fields in various nations, the ranking of crops, and the growth of a specific field's crop and relative yields over the last three years. Developers can be notified by users when crops and fields are defined incorrectly; this information is then utilized to enhance machine learning algorithms. Additionally, the map sheds light on regional and worldwide agricultural production trends.
The precision farming methods in Russia are still in their infancy. For instance, in, staff members of IndorSoft LLC (Tomsk) introduced the IndorAgro software package, which uses technologies from geographic information systems, remote sensing, vegetation indices, three-dimensional modelling, and deep neural networks to support precision farming, efficiency evaluation, and agricultural activity planning. The application creates and evaluates three-dimensional terrain models of agricultural land and processes optical and multispectral aerial photography data. Additionally, it computes vegetation indices and enables automated identification and assessment of depressed, wooded, and other unique zones on agricultural land terrain models. Currently, the company's official website does not provide this software package for sale. It is said on the same page that this group of programs was one of the company's 2018-2019 projects. Building a 3D terrain model from photos taken with industrial unmanned aerial vehicles (UAVs) and automatically recognizing and analyzing a 3D terrain model using convolutional neural networks technology were the foundations of the technology used to gather factual data on the condition of forested areas. Delphi was used in the development of the software package.
The Federal State Budgetary Scientific Institution "Federal Scientific Agro engineering Center VIM" (Moscow) developed the software package which uses soil image analysis to identify weed items. The method of calculating the contour by examining the gray level uses the formulas of the mathematical model of technical vision that the authors previously devised for the calculation. By controlling weeds in the field, the program can reduce the need for herbicides and enhance soil cultivation and crop output, especially for beets. Taking into consideration operations in conditions of fluctuating lighting levels and other field operating parameters of this robotic complex, the application enables you to detect and manage weeds in the early phases of beet growth. Although the writers have several patents, there was little information available regarding the commercialization of their innovations. The authors' articles don't go far enough in describing the mathematical models they employed. In order to promote the sensible use of agricultural land, South Ural State University declared in 2018 that it was developing a "Geo portal system for agriculture monitoring". Even now, development is still going on. A database of agricultural facilities in the Chelyabinsk area is part of the system. The system's primary benefit is that it was built using open components, which guarantees the availability of software tools and the potential to extend functionality independently and with the help of global colleagues. A user-friendly web interface is used to implement the system. This article will examine a yield prediction system that uses the season's average vegetation index value.
Computers can learn without explicit programming thanks to a form of artificial intelligence called machine learning. Reducing the quantity of input variables in a machine-learning model is known as feature selection. The process of giving an input data point a label is known as classification. The process of determining the optimal solution to a problem is known as optimization. Important methods for enhancing the performance of machine learning models include feature selection classification optimization and machine learning. The particular problem being tackled and the resources at hand determine which strategies should be used.

II.THE DESIGN OF THE SYSTEM

The created system functions as Figure 1 illustrates. In order to estimate the projected yield per hectare, this system analyzes the original image of the object of interest, obtains timely vegetation index information, and uses that information.

A. Preliminary information

The GeoTIFF picture is obtained as initial data in the first step. It could be a satellite image or one taken by a UAV's multispectral camera. GeoTIFF is a multilayer raster image with millions of pixels in both situations. One of the bands is represented by each layer in this illustration. The following bands, for instance, are present on the Landsat satellite:

1) Blue (0.45-0.52mkm)
2) Green ( 0.52-0.60mkm)
3) Red (0.63-0.69 mkm)
4) The near-infrared spectrum (0.76-0.90 mkm)
5) Infrared middle (1.55-1.75 mkm)
6) Infrared spectrum (2.08-2.35 mkm)
7) Thermic (10.4-12.4 mkm)
B. Preprocessing Data

The list of bands is the same for every other camera and satellite. Furthermore, the extra layers of such an image allow us to store some more data. We must create the pixel vectors for each band in order to do additional analysis (see picture 2).

C. Segmenting the Field

We can categorize and identify the image data once we have such a vector. The recognition and classification process is used to segment distinct objects of interest and refine the boundaries of the studied areas once we already know the borders of all objects of interest, as they are all saved in the "Geoportal system for agriculture monitoring" database.

Figure 3 provides an example of field segmentation, with T1 representing a wet segment, T2 and T3 representing segments in a forest's shadow, T4 representing a segment devoid of any anomalies, T6 representing a segment close to the road, and T5 representing a segment in a shadow close to the road (with two influencing factors combined).

Every location within each section ought to be found independently. It is helpful to divide all objects of interest into tiles that belong to the designated areas as soon as the pixel count reaches a significant amount. It makes it possible to drastically cut down on the quantity of data being studied. Figure 4 illustrates the data restructuring required to combine the pixels into tiles. Each of the n vectors in this case has a length equal to the number of bands, where n is the number of pixels.

We obtain the same data for the remaining time points of the vegetation period by combining this data into tiles. Figure 5 illustrates the 3-dimensional matrix that we create after compiling the data for the entire vegetation season. Here, Li is the number of a layer (band), Ti is the number of a tile, and tk is the time moment.

D. Calculating the NDVI

We determine the NDVI value for every moment using these vectors. The normalized difference vegetation index, or NDVI for short, is a statistic that uses satellite data to describe an area's vegetation density. The NDVI is computed using the formula below: NIR is the reflection in the near-infrared spectrum (band 5 of a Landsat image); RED is the reflection in the red range of the spectrum (band 4 of a Landsat image); and NDVI is the normalized difference vegetation index. The formula for calculating the NDVI is
NDVI = (NIR-RED)/(NIR+RED) (1)
where RED is the reflection in the red region of the spectrum (band 4 of a Landsat picture); NIR is the reflection in the near-infrared spectrum (band 5 of a Landsat image); and NDVI is the normalized difference vegetation index. Figure 6 illustrates the process for obtaining the NDVI for a given field with known borders saved as a *.shp file.

E. Modelling and Time Series

Consequently, for every tile in question, we get the collection of time series that correspond to the dynamic process. The model of this process is obtained using the sevectors (time series) as initial data. This model can be used for crop forecasting in later vegetative phases. Undoubtedly, external factors like the weather, lunar cycle, humidity, etc., are used to rectify it. However, these factors are not taken into account in this work and are the subject of other, extensive research. Any model that uses the time series as initial data can be used. Either a determinate model utilizing recurrence equations or a known regression or deep learning model may be used.

To implement the studied by [?] techniques to find the coefficients a1, a2, a3..., am ∈ R of an m-th order quasi linear autoregressive model, the quasi-linear n-factor auto regression equation of the m-th order can be expressed generally as follows.

y_(t )=∑_(j=1)^(n(m))▒a_(j ) g_j ({y_(t-k) } _(k=1)^m )+ε_t, t=1.2,…..,T (2)
by current data regarding the values of state variables {y_(t )∈R} _(t=1-m)^Tat instants of time t; Here, gj: ({y_(t-k) } _(k=1)^m ) →R, j=1,2,…, n(m) functions are provided, and {ε_(t ) ∈ R} _(t=1-m)^T are identified mistakes.
The time series {y_(t )∈R} _(t=-1-m)^T is obtained by the GLDM technique of length T + m ≥Using 1 + 3m + m2 as input data, the optimization problem is solved to get the factors a1, a2, a3,..., am ∈ R.

∑_(t=1)^T▒arctan|∑_(j=1)^(n(m))▒a_(j ) g_j ({y_(t-k) } _(k=1)^m )| -y_t→ _(({a_j } _(j-1)^(n(m)) ))^min⊂ R (3)
The distribution of Cauchy
F(ξ) =1/π arctan(ξ)+1/2
possesses the highest entropy of all random variable distributions with no volatility or mathematical expectation. For this reason, function arctan(∗) is used in this study.
Additionally, we examine a model of m-th order with quadratic nonlinearity; thus, the fundamental collection of g(∗) functions comprises

g_((k)) ({y_(t-k) } _(k=1)^m )=y_(t-k,)
g_((kl)) ({y_(t-k) } _(k=1)^m )=y_(t-k ). y_(t-l,)
k = 1,2,……m; l = k, k+1,….., m
Of course, in this instance n(m)=2m+ c_m^2 = m(m+3)/2 and g(∗) functions can have any numbers.
The predictor creates the family of m-th order difference equations indexed by t = 1,2,...,T −1,T.

¯(〖y[t]〗_(τ ) ) = ∑_(j=1)^(n(m))▒〖a_(j ) ^* 〗 g_j ({¯(〖y[t]〗_(τ-k ) )} _(k=1)^m ),
τ = t ,t+1, t+2, t+3, ………., T-1, T, T+1,…… (4)
for lattice functions ¯(y[t)]) with values ¯(〖y[t]〗_τ ) that are interpreted as being constructed at time moment t, the projections for y_τ. Let us use the solution of the Cauchy problem for its difference equation (4) under the initial conditions
¯(〖y[t]〗_(t-1) ) =y_(t-1) ,¯(〖y[t]〗_(t-2) ) = y_(t-2,) ………, ¯(〖y[t]〗_(t-m) ) = y_(t-m,) t=1,2,…….,T-1,T (5) to determine the values of the function ¯(y[t]) . Therefore, the set of potential prediction values of ¯(y_τ ) is represented by y_τ= { ¯(〖y[t]〗_(τ ) )} _(t=1), We also estimate the probabilistic properties of the y_τ value using this set.
GLDM estimation is an example of a concave optimization problem. With the right parameters, GLDM-estimates perform best for probability distributions of errors with heavier (than normal distribution) tails and are resilient to the existence of a connection between values.
Figure 6 provides an example of this model's application, displaying the winter wheat modelling results for the Stavropol region. The average data for a single field is included in this dataset However, if a field is divided into multiple tiles, taking an average for a single tile is not difficult as long as we have a map of the field and its environs. The experiment under consideration began with observations at the end of September, proceeded until the beginning of July when the wheat began to ripen, and concluded with the final autumnal observation at the beginning of December. The next observed value was acquired in February when the snow started to melt. We employ the linear approximation to obtain the missing data. The following model coefficients for the model (see fig. 7) can be obtained from the identification results for this experiment:
y_(t )=3.24249 .y_(t-1)-2.144406.y_(t-2) +
(-6.66915. y_(t-1 )^2-4.8095.y_(t-1 .) y_(t-2 )+11.3432.〖 y〗_(t-2)^2) (6)
The errors are MAE=0.02095274 and MBE=0.01885199, and the loss function value is 0.9172588.

F. Interpretation of the Results

Once we have this model, we can use the first two NDVI values to anticipate the crops for the upcoming season. We may then modify the model while acquiring the real NDVI values. This enables us to predict crops at the start of the growing season. We can utilize methods like and to forecast the crops for the current season. Regarding, as this method is uncommon, it uses data from Spain on crops that are not relevant to Russia. Let's look at the method used for the Voronezh region in . The findings of building the regression dependence of yield U on the values of the integral indicator S(which stands for the mean value of the growing season NDVI index over the interval) are used by the authors. The equation that results is
U= -37.06 + 129.83S. (7)

Crop yields can be predicted using this equation. Assume that the value of S for the example under consideration is equal to 0.508. Then, 29.89 q/ha is the expected value of U (and 37.5 q/ha if we only look at data after February). Therefore, even if this formula shows a stronger correlation between productivity U and the integral indicator S (average NDVI) than with the model's intensive features, it is still applicable for the domains taken into account in . Additional statistical research utilizing the average NDVI for at least 100 fields in a region under study is required if we need to determine the U value for another region of the Russian Federation.

Let's look at the model of linear regression [31].
U= -22.5 + 100.34S. (8)
With the information from [31], we were able to determine our own mode.
U= 14.0277 + 34.8439S. (9)
Table I shows the crops that were calculated using these calculations.
U= 14.0277 + 34.8439S. (10)
Using these formulas to calculate crops is shown in Table I.
The contend that the formula they came up with shows a stronger correlation between yield U and the integral indicator S (average NDVI) than it does with the features of the model they employ. It should be noted, however, that even though the results for the data from another region appear to be satisfactory, the formula from applies to the agricultural lands of the Voronezh region and increases the yield forecasting error in other regions with a different climate. The same can be said for other crops and types. Therefore, it is suggested that after gathering information on the actual values of vegetation indices and crop yields, a method for creating a software module for figuring out the regression model's coefficients be developed in order to obtain a more accurate predictive value of yield U for another crop or another region of the Russian Federation. In this method was examined for a single area on a sample of 105 fields. The verification and adaptation of this model for data from other climatic zones, the pursuit of a general, flexible approach, the automation of the model coefficients, the investigation of the impact of different external factors, as in the approach and the inclusion of economic calculations for estimating and forecasting the unit cost of agricultural crops are thus the focus of additional research.
Future study will therefore focus on validating and adjusting this model for data from the Chelyabinsk region, examining the impact of various external factors such as approach , and include cost-effective computations for crop cost assessment and forecasts.
III. USING IOT AND AI TO IMPROVE AGRICULTURAL FORECASTING
This study suggests a novel strategy to enhance food security and agricultural practices, particularly in Russia. It uses AI and IoT to enhance crop forecasts. Many precision agricultural software programs don't meet Russia's food security requirements or the data accessibility norms of independent professionals. Aerial photography data is collected, processed, and made accessible for study by any analyst using our platform. Our technology uses IoT-enabled drones to capture high-resolution images of farms. Plant height, leaf color, and crop density are among the growth and health trends depicted in these images. After processing this data, AI systems generate vegetation indices for predetermined time periods and construct mathematical models. Future crop yields are accurately predicted by these models. Weather, soil, and yield data are used in forecasts. Additionally, it forecasts crop prices by taking into account production costs, market trends, and yield. Although promising, our solution could be enhanced. Future initiatives include enhancing IoT sensor settings, expanding data sources, and improving AI algorithms to increase forecast accuracy. Drone photography and satellite images could be used to evaluate crop conditions. Enhancing data management is an additional area of focus for research. The system generates a lot of data, thus data processing, retrieval, and storage must be effective. Research could use contemporary database technologies or cloud computing to optimize these procedures. Lastly, our strategy enhances crop forecasts and food security by utilizing IoT and AI. With trustworthy, easily accessible data for experts and decision-makers, precision farming progresses. Through research and system improvement, we hope to advance efficient, sustainable agriculture.
, Claims:1. The outcomes of the aerial photography data gathering and processing system are encouraging for the crop forecasting system. It computes vegetation indices for a specific date using a dependable and efficient data display method.
2. Crop forecasting systems use a dependable and efficient data presentation technique to calculate vegetation indices for a given date. They also provide methods for forecasting a field's yield in the upcoming season and modelling the vegetation process using high-quality determinate mathematical models.

Documents