Dengkou County, located in Bayannur City, Inner Mongolia Autonomous Region, China, was selected as the study area ( Figure 1 ). The highest altitude of Dengkou County is approximately 1951 m, with a total area of approximately 4,200 km². The climate is a temperate continental monsoon type, with an average annual temperature of approximately 9.16°C. The average annual rainfall is 106–155 mm. The temperature difference between day and night is large, approximately 14.5°C, and the frost-free period lasts for 136–144 days. From the remote sensing image shown in Figure 1 , Dengkou County is located in the upper reaches of the river bend irrigation area and at the eastern edge of the Ulanbuh Desert in the southwestern Inner Mongolia Autonomous Region. Dengkou County is one of the nine experimental demonstration counties for sand control and prevention in China because of its diverse topography, which is dominated by sands, mountains, and plains, and is known as “seven sands, two mountains, and one plain”. The topography and climate of Dengkou County provide a unique foundation for the development of modern agriculture and animal husbandry. In addition to planting traditional crops such as Avena sativa L., Helianthus annuus L., Zea mays L. and Glycine max (L.) Merr., medicinal plants such as Cistanche deserticola Ma, G. uralensis, and Astragalus membranaceus (Fisch.) Bunge is grown in Dengkou County, and the area under cultivation is increasing annually (Jia et al., 2022; Le and Lihua, 2022).

The primary data source for this study was the Gaofen-1 Satellite Wide Field of View (GF-1 WFV) images, which can be obtained free of charge from the Land Observation Satellite Data Service (LOSD) of the China Centre for Resources Satellite Data and Application (CRSDA). Application (https://data.cresda.cn/#/home) of the CRSDA. The technical specifications of its payload are shown in Table 1 . The GF-1 satellite was not only launched earlier, but also has a higher spatial resolution, a larger width and a shorter revisit period (Tong et al., 2023; Yuan et al., 2024). Therefore, more historical image data can be acquired, the details are more clearly shown, more information is covered, and more image data can be acquired within the same period.

For the acquired GF-1 WFV images, radiometric calibration was first performed using the GF-1 absolute radiometric calibration coefficients released by CRASAC for the corresponding period. Next, the GF-1 WFV spectral response function was imported into the spectral library file, and atmospheric correction was applied using the FLAASH module. Finally, ortho-rectification was conducted with the GMTED2010 elevation data. After radiometric calibration, atmospheric correction, and ortho-rectification, the data significantly reduce atmospheric influence and achieve high geometric accuracy, allowing for better stacking of multi-period images.

After ensuring the authenticity and the accuracy of the surface reflectance data by preprocessing the remote sensing images and eliminating the noise, distortion, and interference factors in the images, the surface reflectance values of the main feature types in Dengkou County in 2022 were extracted using field survey data and remote sensing image data, and the results are shown in Figure 4 .

From Figure 4A , the January–December surface reflectance of each feature type in the blue band was similar, and except for water bodies and mountains, the curves of all of the other feature types were consistent with each other, showing a decreasing trend in spring (March-May), increasing in summer (June–August), decreasing and then increasing in autumn (September–November), and increasing and then decreasing in winter (December–January). In autumn (September–November), the curves first decreased and then increased, whereas in winter (December–January), they first increased and then decreased. The G. uralensis reflectance curves distinctly differed from the other feature types in January, February, September, and October. As shown in Figure 4B , the surface reflectance for each feature type under the green band exhibits distinct curve characteristics in spring, summer, and autumn, except for the overall trend of increasing and then decreasing in winter. Likewise, the G. uralensis reflectance curves were distinguished from other feature types in January, February, September, and October. In Figure 4C , the surface reflectance curves for each feature type under the red band are similar to those under the green band and show an overall increasing and then decreasing trend in winter. The G. uralensis reflectance curves were distinguished from the other feature types in September. From Figure 4D , all of the feature types under the Near Infrared (NIR) band showed the same pattern of increasing and then decreasing, except in winter, with peaks of different sizes from June to October, which is similar to the reflectance curves of all feature types under the blue band. The G. uralensis reflectance curves differed significantly from those of the other feature types in June. Therefore, calculating the difference between the blue and near-infrared bands can enhance the distinction between spectral curves, facilitating the identification and extraction of G. uralensis and other feature types.

As shown in Figure 5 , certain feature types can be accurately recognized under month-specific imagery, and a collation of such feature types and the months in which they are recognized are as follows: buildings (September), water bodies (January–December), deserts (August–September), mountains (February, September), oats 1 (September), wetlands (March-May), and unknown 2 (April–May). In addition, the feature types exhibit similar curve changes for the same remote-sensing image. Except for the imagery from June and August, the reflectance values of each band in the remaining images show a trend of blue, green, red, and near-infrared bands from small to large. Therefore, it is difficult to accurately identify G. uralensis and other features using remote sensing images at certain times. Moreover, according to the reflectance values of the four bands for the main feature types, a result consistent with the analysis in Figure 4 can be obtained; that is, the calculation between the blue and near-infrared bands is more conducive to increasing the differences between G. uralensis and the feature types, thus facilitating the identification of each feature type.

Based on these results, the band importance was evaluated using RF, and the results are shown in Figure 6 . Figure 6A shows that the raw importance contribution rate is highest for the NIR band, followed by the blue band with a higher contribution rate, and the red and green bands with the lowest raw importance contribution rate. The normalized importance contribution rate was highest for the blue band, followed by the near-infrared band, and lowest for the red and green bands. After further organizing the band contribution rates by band types, Figure 6B shows that the raw contribution rates ranked from largest to smallest as NIR, blue, red, and green bands. The normalized importance contribution rates from the largest to smallest were the blue, near-infrared, red, and green bands. This result is consistent with the aforementioned results and shows that combining the January–December bands into a model enhances the classification of each object type.

Based on the result that the calculation between the blue and near-infrared bands is more conducive to increasing the difference between G. uralensis and other feature types obtained from the analysis of spectral characteristics of feature types, the difference between the near-infrared band and blue band was increased using the formula (A−B)/(A + B) in combination with commonly used vegetation indices formulas (Zhai et al., 2023; Pei et al., 2024; Tang et al., 2024; Zhen et al., 2024). It was named GUVI, and the formula is shown in Equation (6) .

where NIR represents the near-infrared band, and B denotes the blue band.

To verify the reliability of the GUVI for identifying the distribution area of G. uralensis, seven commonly used vegetation indices ( Table 2 ) were selected to construct the G. uralensis identification model for comparison and verification (Giovos et al., 2021; Zeng et al., 2022; Han et al., 2024). Because it is difficult to accurately identify and extract the distribution area of G. uralensis and other feature types under a single remote sensing image, and according to the results of the band importance contribution rate, this study, the monthly GUVI, and seven other vegetation indices were calculated separately after processing the remote sensing images from January to December 2022 and were combined into eight long time series data as feature sets in the order of months. Based on the sampling point data of the main feature types in 2022, 70% were selected as the training set and the remaining 30% as the validation set. Eight identification models of G. uralensis were constructed using the RF algorithm, and the reliability of the GUVI was verified by comparing the classification accuracies of the RF algorithm for the eight models. The classification effect of different classification algorithms on the GUVI feature set construction was compared by comparing the classification accuracy of the RF classification model constructed with the GUVI feature set and the SVM classification model (Ko et al., 2024; Yang et al., 2024). The results are shown in Table 3 .

Table 3 shows that the model constructed with DVI as a feature had the highest OA and KC of 93.79% and 92.29%, respectively, and its G. uralensis UA and PA were 93.68% and 97.22%, respectively. The model constructed with GNDVI as a feature had the highest G. uralensis UA of 97.13%, OA of 92.10%, KC of 90.23%, and G. uralensis PA of 98.02%. The model constructed using GUVI as a feature had the highest G. uralensis PA of 97.26%, OA of 93%, KC of 91.38%, and G. uralensis UA of 97.32%. The model constructed with GUVI as a feature was better than the DVI in terms of G. uralensis UA and PA values; it was better than the GNDVI in terms of OA, KC, and G. uralensis PA values; OA and KC were not the highest, but the performance was excellent. In addition, through the classification accuracy of the two classification algorithms, SVM and RF, for the GUVI feature set that RF has a better classification effect, which is higher than the classification accuracy of SVM in the values of OA, KC, G. uralensis UA and G. uralensis PA. From the vegetation indices feature set model classification results graph ( Figure 7 ), the GUVI_RF classification results were more accurate than the other results and were able to identify the subtle differences between the habitat types more accurately, producing fewer misclassifications. Even using different classification algorithms to classify the GUVI feature set yielded better results, indicating that the method is more robust and efficient in recognizing G. uralensis. That is, the classification of the GUVI model with RF can effectively recognize the distribution area of G. uralensis in Dengkou County, and it has certain applicability in other G. uralensis planting areas.

According to the results discussed in Section 3.2, the model of the GUVI combination for January–December 2022 can be used as a feature set to effectively identify and classify G. uralensis and other feature types using the RF algorithm. Therefore, this method can identify and classify G. uralensis and other feature types in historical remote sensing images from 2014 to 2021. However, owing to the lack of actual sampling point data for past feature types, the corresponding training and validation sets could not be selected. Therefore, in this study, the spectral curves of the main feature types and different growth years of G. uralensis were obtained based on the average values of the training and validation sets under the GUVI combination model for January–December 2022 to analyze the differences between the spectral curves of G. uralensis and other feature types, while simultaneously understanding the difference in the spectral curve of G. uralensis with different growth years. This enabled more comprehensive information on G. uralensis to be captured while distinguishing G. uralensis from other feature types without actual sampling points, allowing the RF algorithm to identify historical G. uralensis distribution areas more accurately.

The spectral curves of the main feature types under the GUVI model for 2022 are shown in Figure 8 , with different GUVI values and spectral curves for each feature type. In terms of the GUVI values, the G. uralensis spectral curve was significantly different from the other feature types for eight months (January, February, May, June, July, September, November, and December). The feature types with GUVI values similar to those of G. uralensis in the remaining months were oats in March, unknown 1 in January and April, oats 2 and maize in August, and wetlands in October. These feature types in other months are significantly different from those of G. uralensis and can be distinguished based on the values for other months. Based on the curve analysis, the feature type that has similar characteristics to the G. uralensis curve is wetland, and the curves are all characterized by a rise in spring, a peak in July, a low peak in October, and a trough in winter in January of the following year. However, both had similar values, except in October, and the values of G. uralensis were greater than those of the wetland in the remaining months. Therefore, according to Figure 8 , it can be concluded that the spectral curve characteristics of G. uralensis under the GUVI model are different from those of other feature types. Thus, the training and validation sets can be selected under the GUVI model based on the curve characteristics, and G. uralensis can be identified by the RF algorithm from 2014 to 2021.

G. uralensis is a perennial herb or half-shrub, and its growing status has been shown to vary significantly from year to year (Xi, 2020), and the reflected spectral features will be different. In order to more accurately identify and extract the distribution areas of G. uralensis in historical remote sensing images, this study further compiled the characteristics of the spectral curves of G. uralensis in different years under the GUVI model for 2022, based on the G. uralensis year information provided by local growers at the time of sample collection.

As shown in Figure 9 , the characteristics of the spectral curves for different years of G. uralensis were different. In terms of values, annual G. uralensis had lower values than biennial and triennial G. uralensis in all months except June and October. Biennial G. uralensis had higher values than triennial G. uralensis in January, May, June, July, and November, while triennial G. uralensis had higher values than biennial G. uralensis in the remaining months. The spectral curve of annual G. uralensis started to rise in March, surged in April–June, peaked in June, and then started to decline, and was more moderate in July–August, declined in September, and then started to rise in October, declined again in November, and rose again in December. The spectral curve of biennial G. uralensis, however, started to fall in January, began to rise after March, peaked in July, and then fell to its lowest value in October, after which it started to rise and became relatively flat in November and December. The triennial G. uralensis curve rose in January, declined throughout February, was relatively flat in March and April, rose again, reached a high value in July, remained relatively flat in August, and then declined until it began to rise again after November. Therefore, the spectral curve characteristics of G. uralensis across different growth years should be comprehensively considered when selecting sample points from historical remote sensing images to enhance the RF algorithm’s accuracy in identifying G. uralensis distribution areas.

According to the results in Section 3.3, under the GUVI model, the spectral curves of different feature types and different growth years of G. uralensis have their characteristics, which can be used to select samples of the main feature types in Dengkou County from historical images. Therefore, after preprocessing the remote sensing images from 2014 to 2021, the monthly image GUVI was calculated, and the different years GUVI were combined into eight feature sets of the RF algorithm in the order of January–December. The features were deciphered according to the spectral curves of G. uralensis with different growth years as well as other feature types. The visual interpretation was used to obtain the sample points of G. uralensis and the other feature types in the historical images from 2014 to 2021. After selecting the training set and validation set in the ratio of 7:3, the distribution area of G. uralensis was identified using the RF algorithm, and its accuracy was verified.

The classification accuracies of G. uralensis distribution areas of RF classification model constructed with GUVI feature set for 2014–2021 are shown in Table 4 . Among them, the lowest value of OA was 89.72%, the lowest value of the KC lineage was 87.16%, the lowest value of G. uralensis UA was 69.47%, and the lowest value of G. uralensis PA was 71.08%. Therefore, the overall classification accuracy and individual classification accuracy of G. uralensis were good for 2014–2022 and could be used for G. uralensis range extraction.

The spatial distribution and changes in the G. uralensis distribution area from 2014 to 2022, extracted from the identification results of the G. uralensis distribution area using the RF algorithm under the GUVI model, are shown in Figures 10 , 11 . As shown in Figure 10 , the distribution of G. uralensis gradually expanded to the central region from its initial main concentration in the northeast of Dengkou County and is now distributed in the southwest. The distribution area gradually changed from small and scattered at the beginning to a large area of concentrated distribution. Figure 11 shows the changing trend of the G. uralensis area, which first increased and then decreased. The most significant increase in the G. uralensis area was observed in 2018. As G. uralensis is a perennial plant, it is difficult to systematically count the area of G. uralensis planting and harvesting each year; however, it generally shows a stable planting pattern. In addition, although Dengkou County has a long history of G. uralensis planting, the planting area was small before 2018, and only after 2018 did it begin planting on a large scale.

Although multispectral remote sensing images offer advantages such as wide coverage and high spatial-temporal resolution, their limited spectral bands can pose challenges for accurate classification. When relying on original image spectra to classify and identify feature types, often leads to suboptimal accuracy. In addition, the fragmentation of agricultural land systems in terms of cultivation type and area, as well as the influence of the phenomenon of “different spectra for the same object and the same spectra for the same object” and the mixed pixel effect, making remote sensing identification and extraction of crops more complicated than that of the natural vegetation (such as woodland and grassland). Zheng et al. (2022) investigated remote sensing identification of film-covered farmland in the Loess Plateau using various combinations of spectral, exponential, and texture features. They found that combining spectral, exponential, and optimal texture features yielded the best classification results. Liang et al. (2021) successfully constructed remote sensing features for citrus by combining spectral, climatic, textural, and spectral indices and reported that the normalized vegetation index (NDVI) was particularly outstanding in the identification of citrus. Liu et al. (2023) used Google Earth Engine, Sentinel - 2 data and winter wheat phenology as prior knowledge to develop new indices, including the Normalized Differential Phenology Index (NDPI), Wheat Phenology Differential Index (WPDI), Normalized Differential Wheat Phenology Index (NDWPI), and Plastic Mulched Index (PMI). This method is used to obtain the spatial distribution and planting area information of winter wheat. The accuracy of the proposed method is higher in the early wintering stage and the regreening stage, which are 82.64% and 88.76%, respectively, and has strong spatial and temporal transfer. In order to improve the distribution area of a certain plant, it is necessary to find the differences with other plants by understanding the growth characteristics, spectral characteristics, temporal phase changes, and other aspects of information, and then increase the differences through the method of band operation, and further increase the differences by combining the features to construct a more accurate and effective remote sensing identification model. This would improve the identification accuracy and classification effect of the distribution area of the plant. In the classification algorithms, although the principles of the different algorithms are different, the classification accuracy of the two different algorithms applied in this study for the GUVI feature set shows that this classification model has a certain degree of robustness and the efficiency of recognizing G. uralensis is high, thus indicating that the RF classification model constructed based on the GUVI feature set is not overly influenced by the characteristics of a single algorithm and can classify G. uralensis stably in the framework of different algorithms. In addition, the classification accuracy of G. uralensis distribution areas under the RF classification model based on the GUVI feature set across different growth years shows that the model can maintain relative stability in the process of year change, indicating that the RF classification model based on the GUVI feature set has the common characteristics of G. uralensis distribution areas at different times and is not subject to the excessive interference of factors such as environmental changes between years; therefore, it has good portability. Given that the sampling points in this study covered a variety of G. uralensis growth conditions ranging from good to poor growth ( Figure 2 ), which ensured the diversity and representativeness of the data, the RF classification model constructed based on the GUVI feature set was considered to have good generalizability. However, due to the influence of many natural and human factors such as region, climate, soil, farmers’ planting methods and cycles, the universality of the RF classification model based on G. uralensis feature set in different regions may be limited, and it is necessary to adjust the parameters and optimize the model for specific regions. In practical applications, it is necessary to fully consider its limitations and combine other methods to improve the accuracy and efficiency of classification.

There are obvious differences in the spectral characteristics of different plants due to the differences in leaf tissue structure and pigments. For the same plant, the spectral characteristics may change significantly at different growth stages or under different environmental conditions. As G. uralensis is a perennial herb or half-shrub, understanding its spectral characteristics at different ages forms the foundation for classifying feature types over an extended time series. The current study demonstrated that the spectral curve characteristics of different annual G. uralensis varieties were significantly different. This may be because one-year-old G. uralensis is in the early stage of growth, its leaves are not yet mature, and the content of chlorophyll and other pigments is relatively low. Meanwhile, its biomass is small, the leaves are thin, and the organizational structure is not close enough, which makes its absorption and reflection of light weaker, thus exhibiting lower characteristics in the spectrum. In contrast, the spectral curves of biennial and triennial sweetgrasses were more similar because of the maturity of the leaves, increased content of chlorophyll and other pigments, accumulation of more biomass, and other reasons. However, there may be a few differences owing to differences in accumulation, environmental factors, or the management styles of farmers, which need to be further analyzed. From the overall analysis of these three curves, reveals a clear peak from May to August, a trough from March to April and September to November. This trend closely correlates to the growth cycle of G. uralensis. From March to April, G. uralensis recovers from dormancy, the leaves are not yet fully expanded, and the chlorophyll content is low; therefore, the spectral characteristics are not obvious and show a trough. From June to August, this is the most vigorous growth period of G. uralensis, with high chlorophyll content, lush leaves, significant reflection and absorption spectrum characteristics, forming a peak. From September to November, G. uralensis enters the late stage of growth, leaves begin to senesce, chlorophyll gradually decomposes, and the spectral characteristics are weakened, again forming a trough. From December to February, the ground surface is covered with snow, which has high reflectivity, making the spectral features appear higher than from March to April and September to November. This phenomenon further enriches our understanding of the variation in the spectral characteristics of G. uralensis and provides a useful reference for the subsequent classification of feature types.

The distribution area of G. uralensis in Dengkou County has experienced significant changes from 2014 to 2022. Studies during this period showed that the distribution area of G. uralensis not only expanded, but also its concentrated distribution became more prominent. Initially, G. uralensis was mainly distributed in a small area in the northeastern part of Dengkou County, but over time, it gradually expanded to the central and southwestern parts of the county, eventually forming a large area of concentrated distribution. This change reflects the importance of G. uralensis in the region and shows the complex interaction between natural conditions and human factors. First, the geographical conditions in the northeastern part of Dengkou County provide a favorable environment for G. uralensis growth. This region benefits from the Yellow River and has abundant water resources, creating an environment suitable for G. uralensis growth. The abundance of water resources enables G. uralensis to obtain the water required during the growing season, thereby improving its growth rate and quality. In addition to the influence of natural conditions, the local agricultural technology extension departments play a key role. With the spread and popularization of G. uralensis cultivation techniques, farmers’ cultivation and management skills have significantly improved. These departments may help farmers master modern cultivation methods and management techniques by providing training and guidance that have improved the overall efficiency and yield of G. uralensis cultivation. In addition, with the large-scale development of the G. uralensis cultivation sector, cooperation and exchanges among growers have become increasingly frequent. Growers have further improved the efficiency and competitiveness of G. uralensis cultivation by sharing cultivation experiences and optimizing cultivation patterns. Such cooperation enhances ties between growers and provides a good basis for the intensive production of G. uralensis and promotes its distribution on a wider scale. However, the areas planted with G. uralensis fluctuated significantly between 2014 and 2022. This fluctuation may be due to the combined effect of many factors. For example, inconsistencies in growers’ harvest years may lead to changes in the areas under G. uralensis cultivation. Due to the fluctuations in market demand, growers may adjust the planting area of G. uralensis according to the change in market price and demand in each growing season. In addition, the intensification of market competition may impact the decision-making of growers. A few growers may choose to reduce the area of G. uralensis cultivation because of competition from other crops, whereas others may expand their cultivation because of favorable market prospects. This fluctuation in planting area reflects the complex relationship between the market environment and farmers’ decision-making and indicates a certain degree of vulnerability in the G. uralensis farming industry. In summary, the change in the G. uralensis distribution area in Dengkou County resulted from the combined action of natural conditions and human factors. In the future, in order to promote the sustainable development of G. uralensis plantations, it is necessary to continue to pay attention to and study the growth characteristics of G. uralensis and market demand and to seek reasonable planting strategies to achieve sustainable development.