Preliminary results of automatic cotton crops mapping using remote sensing data

. The paper presents the results of application of the method of automatic generation of representative and unbiased set for in-season cotton crop mapping, based on crop simulation model, previously parameterized using ground truth and satellite data. The method provided confident mapping of cotton fields without using actual ground-truth information or a-priori information about their in-season phenology. Overall mapping accuracy calculated using relevant ground truth data for cotton fields has reached 95.6 %. Consideration of time series of NDVI values as a model of phase characteristics allowed using relatively simple criteria to identify typical representatives of the selected crop on the basis of analysis of their seasonal phenology and made it possible to build a reference sample for modeling and further classification.


Introduction
The cotton sector plays an important role in the economy of the Republic of Uzbekistan.The reforms implemented by the government in the cotton sector have become an important element of the country's planned development and its transition to a market economy" [1].
The Republic of Uzbekistan has implemented comprehensive large-scale measures to improve efficiency of the production process of seed cotton and to introduce highly effective technological process management systems that improve properties of cotton products" [2].
Timely information on crop mapping is of paramount importance for operational assessment of crop condition, crop rotation control and yield forecasting.The use of remote sensing data has obvious advantages, over traditional ground inventory methods, due to the prompt receipt of information, as well as the objectivity and spatial detail of assessments.Automatic methods of processing and analysis of satellite data allow minimizing material and labor costs, ensuring independence, timeliness and repeatability of results.At the same time, this method is of particular relevance for the mapping of large areas of agricultural land in cotton-textile and agro-clusters in the Republic of Uzbekistan [3].
In particular, the largest cotton and textile cluster in the Republic of Uzbekistan -"Buxoro Agroklaster" LLC, has about 100 thousand hectares of acreage, 65,000 hectares of which are located in the Bukhara region.In 2022 season, about 41,350 hectares of these areas, located in the Bukhara region, were allocated for cotton, 22,500 hectares for winter wheat, the remaining crops accounted for about 1,150 hectares [3].At the same time, for the cottontextile clusters, first of all, it is important to have operational data on the cultivation of the main raw material -cotton, both on their own lands and on contractual farmers lands.In this regard, during the implementation of the work the special attention paid to the possibility of cotton crops mapping.
To ensure the effective management of agricultural production processes, and, in particular, mapping of crop areas in huge areas, it is not possible without the use of automated remote sensing programs.Taking into account the above stated, since 2020 "Buxoro Agroklaster" LLC with the assistance of "Cotton Research and Innovation Center" LLC within the framework of a state grant from the Ministry of Higher Education, Science and Innovative Development of the Republic of Uzbekistan has been developing a remote sensing system for cotton, wheat and other agricultural crops.The "Agro Smart Map Uz" software package is being implemented [4].
The purpose of this software package is to develop and implement a new digital remote monitoring system for generating primary accounting data based on the digitalization of the agricultural sector, automation of accounting processes, which together will reflect agricultural activities in such aspects as an inventory of agricultural land with the creation of a map of fields and crop rotations, agrochemical (agricultural chemical investigation) survey and monitoring of the green mass index (NDVI), agroecological survey (Scouting), analysis of weather conditions (Meteo), precision farming with differential application of seed, mineral fertilizers, plant protection products ( PPP), etc., and also monitoring the movement of equipment, planning and auditing the fact of agrotechnical measures with the formation of analytical data [5].
The scientific significance of the results of the ongoing research lies in the development of a single web platform that will allow, on the basis of information coming from remote sensing modules, stationary and mobile devices, to form historical databases for each field on the readings of weather stations, annual crop rotations , NDVI indices and plant development, the condition of the soil and its fertilization with nutrients, the movement of equipment and material resources, the planned and actually completed field work.The specified platform will also be equipped with a module, for the first time in practice, capable of generating statistical data in the context of administrative-territorial divisions (ATD: region, district, settlement), agricultural enterprises and farms [5].
Modern methods of remote in-season crop mapping often involve the use of a-priori information about the timing of their sowing, stages of development and harvesting [6,7,8], or use stable differences in seasonal dynamics of remotely measured plant characteristics [9].
Almost always, the advantages of time series of satellite images that have proven to be effective for solving these problems" are used [6,7,8,9,10].
The use of a training sample describing the spatial and thematic variability of the characteristics of objects on the Earth's surface is a necessary condition for ensuring the required level of reliability of their mapping based on parametric and nonparametric classifiers.Obtaining timely and spatially distributed information on the crops mapping from independent sources over large areas, as well as the accumulation of such data based on ground surveys or expert analysis of satellite images, is associated with significant organizational difficulties, financial and time costs and is difficult to accomplish.
The paper presents a method for automatic generation representative and unbiased training set for cotton crop mapping, based on crop simulation model, previously parameterized using ground truth and satellite data for 2022 crop season.The method provided confident mapping of cotton fields without using actual ground-truth information or a-priori information about their in-season phenology.

Materials and methods
Ground data were presented by the results of ground surveys of 108 fields of "Garden Buxoro Agroklaster" LLC (Figure 1), covering the 2022 season.Data included, in addition to the boundaries of fields, the names of crops, the dates of the onset of the main phonological phases and the dates of application of mineral or organic fertilizers, soil type and other data.
Bukhara region is a predominantly agricultural region, where almost all the most important crops in terms of national gross harvest volumes are represented.The territory is "homogeneous from the point of view of GAES global agrostratification" (Fischer et al., 2012), which means minimal differences in soil and climatic conditions and agricultural practices in the study region.Two crops mainly prevailed in the studied fields: of which 69 contours were occupied by cotton and 13 by winter wheat (Figure 1).In other areas studied, the fields were divided as follows: 12 fields under pairs, 2 fields under corn, 2 fields under melon, 1 field under potatoes, 1 field under soybeans and 6 fields were divided under orchards (apple trees, apricot).These cultures were not considered due to their insignificant number.Thus, two crops were studied in the work: cotton and winter wheat, for which the total area of crops is more than 90% of the entire studied sown area of the region.
Seasonal series of NDVI parameters from Sentinel-2A/B (MSI) high spatial resolution satellite images were used to crops mapping based on the created sample as mapping signs.
Seasonal time series of NDVI parameters for cotton and winter wheat (excluding the placement of repeat crops) presented in Figure 2. It is clear from the presented data that the parameters of normalized relative vegetation index of cotton and winter wheat differ significantly from each other.This circumstance allows providing mapping of crops by crop-specific models that include ranges of maximum and minimum NDVI parameters.
The model has been parameterized using ground and remote sensing information for 2022 season for 108 fields within a single farm for cotton and wheat.To assess the capability of the proposed approach, the models used only within the boundaries of the surveyed fields with ground-based crop information, where a comparison between the classes of the generated sample and the crop was possible.
At the next stage, when the model parameters for each of the considered crops were established, it became possible to simulate the process of phenological development of plant structural units and their accumulation of green biomass, taking into account the agrometeorological features of the current growing season.At the same time, it is necessary to have information about sowing dates of modeled crops.To establish the belonging of the current object of agricultural vegetation to one of the classes of modeled crops, NDVI measurements of a particular field and its corresponding in time series of predicted model values were compared.Examples of correspondence of different seasonal time series of satellite and model NDVI values for five randomly selected cotton fields presented in Figure 3.

Model-based mapping results
Regional crop mapping was performed based on the reference set obtained and described above, as well as seasonal time series of multispectral, high spatial resolution Sentinel-2A multispectral satellite images (MSI).The satellite dataset used contained measurements of plant NDVI, which used to estimate green plant biomass and which is quite informative in crop mapping.
The overall mapping accuracy calculated based on 69 fields occupied under cotton.During mapping, 66 fields were recognized without errors, 1 field was not recognized and 2 fields were mistakenly recognized (Table 1).

Discussion
Analysis of the mapping results shows that the proposed method, taking into account the predominance of cotton and wheat crop areas, can be successfully used to in-season mapping of agricultural crops based on satellite data and simulation of seasonal plant phenology.Small errors occur and become larger where there are sown areas of crops that have similar vegetative phenology.However, the absolute values of NDVI indexes of the main crops -cotton and winter wheat -are markedly different (see Figure 2), which provides the possibility of their successful mapping by using of additional parameters to improve the method in the traditional conditions of predominance of wheat and cotton crops.With a hypothetical complete separation of these classes, the overall mapping accuracy could increase up to 99%.
It should be noted that when using the mapping of a particular crop, the phases of vegetative development are also an important factor.For example: during mapping of Upland cotton varieties (Gossypium hirsutum), the vegetation phases may be early compared to the Pima cotton (Gossypium barbadense), which in turn may lead to some deviations or distortions.At the same time, the higher accuracy of crop mapping with reference to the phenology phases gives more reliable crop mapping.This fact will allow to promptly collecting information from large areas under crops.
The most demanded is obtaining operational information on the in-season crop mapping.In this regard, it is promising to study the possibility of early construction of a reference set taking into account the vegetation phases.
Additional efforts are planned to be directed to changing the working conditions of the method in order to provide more rapid assessments, as well as to investigate the impact of these changes on the quality of the model and the mapping accuracy.Despite the fact that the study region is homogeneous from the point of view of global agrostratification of GAES, further work on large heterogeneous territories will require localized parameterization of the model, carried out using local ground survey data or expert interpretation of high spatial resolution satellite images and other auxiliary data.In addition, for the automatic application of the proposed approach in large areas, reliable methods of independent and timely determination of sowing and harvesting dates are needed, which can be based on the use of satellite, meteorological and model indicators or their combinations.

Conclusion
Due to its relative simplicity and versatility, the proposed method may be promising for the development of low-cost and operational technologies for in-season crops mapping based on remotely sensed data, including over large areas.Its application requires a limited set of remote and ground data necessary at the stages of simulation model tuning.Consideration of time series of model NDVI values as benchmarks of phase characteristics allows us to use simple criteria to map typical representatives of selected crops based on the analysis of seasonal dynamics of spectral-reflectance characteristics and to construct a sufficiently accurate reference set for further mapping.This solution can be universal in mapping a wide range of crops in large areas, where the main problem is the impossibility of timely and simultaneous acquisition of reference data of the current growing season by other methods.
We would also like to note that operational crops mapping can contribute to the effective use of the "Forecast of food balance of the country", which will allow to adjust the development strategy of "Food Security" of the Republic of Uzbekistan, while increasing the export-import potential of the country.

Fig. 1 .
Fig. 1.The research region, the boundaries of the fields with ground data and the location of crops in 2022 season.

Fig. 2 .
Fig. 2. Examples of the values predicted by the model and the NDVI values obtained on the basis of remote observations for the 2022 season.

Fig. 3 .
Fig. 3. Examples of correspondence of in-seasonal time series of satellite and model NDVI values for five randomly selected cotton fields.
The research was carried out under the grant of the Ministry of Higher Education, Science and Innovative Development of the Republic of Uzbekistan No. IZ-202010156 "Development of remote sensing system for cotton, wheat and other agricultural crops" with the use of resources of the Center for Digitalization of Agriculture under the Ministry of Agriculture of the Republic of Uzbekistan.

Table 1 .
Matrix of classification results errors based on the reference set obtained from the cotton model.