Disturbed boundaries extraction in coal–grain overlap areas with high groundwater levels using UAV-based visible and multispectral imagery

With high groundwater levels, coal–grain overlap areas (CGOAs) are vulnerable to subsidence and water logging during mining activities, thereby impacting crop yields adversely. Such damage requires full reports of disturbed boundaries for agricultural reimbursement and ongoing reclamation, but because direct measurements are difficult in such cases because of vast unreachable areas, it is necessary to be able to identify out-of-production boundaries (OBs) and reduced-production boundaries (RBs) in the corresponding region. In this study, an OB was extracted by setting a threshold via the characteristics of the cultivated-land elevation based on a digital surface model and a digital orthophoto map generated using an unmanned aerial vehicle (UAV). Meanwhile, the above-ground biomass (AGB), the soil plant analysis development (SPAD) value of chlorophyll contents, and leaf area index (LAI) were used to select the appropriate vegetation indices (VIs) to produce a reduced-production map (RM) based on power regression (PR), exponential regression (ER), multiple linear regression (MR), and random forest (RF) algorithms. Finally, an improved Otsu segmentation algorithm was used to extract mild and severe RBs. The results showed the following. (1) Crop growth heights in a typical ponding basin of the CGOA rendered a fast and efficient approach to distinguishing the OB. (2) In subsequent sample modeling, the red-edge microwave VI (MVIredge), the normalized difference VI (NDVI), and the red-edge modified simple ratio index (MSRredge) combined with RF were shown to be optimal estimators for AGB (R2 = 0.83, RMSE = 0.114 kg·m−2); the red-edge NDVI (NDVIredge), the green NDVI (GNDVI), and the red-edge chlorophyll index (CIredge) acted as strong tools in SPAD prediction using RF (R2 = 0.83, RMSE = 0.152 SPAD); the red-edge modified simple ratio index (MSRredge), the GNDVI, and the green chlorophyll index (CIgreen) via MR were more accurate when conducting the inversion of LAI (R2 = 0.88, RMSE = 1.070). (3) With the improved Otsu algorithm, multiple degrees of RB extraction can be achieved in RM. This study provides reference methods and theoretical support for determining disturbed boundaries in CGOAs with high groundwater levels for further agricultural compensation and reclamation processes.


Introduction
China is one of the largest coal producers and processors in the world, and as an indispensable pillar industry of China, coal mining is a major part of the country's energy and economy. Nevertheless, large-scale coal mining inevitably brings significant landscape damage and detrimental environmental pollution, which now poses a looming threat to the well-being of all mankind (Bell et al. 2000;Bian et al. 2009).
In coal mines with high groundwater levels, the groundwater tends to rise upward easily after subsidence, thus forming a unique water-land compound ecological environment (Ren et al. 2020), and this outcome is most obvious for those mining areas that overlap cultivated land. A coal-grain overlap area (CGOA) is one where there is underground mining and further processing and meanwhile agricultural production is ongoing above ground ), in which case crops are soaked in stagnant water for a long time, resulting in dramatic yield losses. According to regulations pertaining to agricultural reimbursement and reclamation, the final decision about crop losses is affected by the scope of disturbed boundaries (Xu et al. 2014;Hu and Li 2019). However, because of the lack of policies regarding reporting the loss of cultivated land in CGOAs with high groundwater levels, the degree of damage is difficult to describe quantitatively and there is still no consensus on criteria for determining disturbed boundaries and the corresponding compensation levels. Therefore, the precise extraction of disturbed boundaries in subsidence areas is a prerequisite as well as an important basis for realizing reimbursement and reclamation.
Currently, China has no unified standard for defining the boundaries of damaged cultivated land. It is ambiguous to identify disturbed boundaries via the most frequently used 10-mm subsidence contour, this being because most of the cultivated land is only slightly affected. In such cases, this blurred definition leads to unnecessary increases in project budgets, thereby placing extra burdens on national economic expenditure and wasting valuable resources. To address this uncertainty, many scholars have carried out studies based on geological information to simulate surface subsidence Hu and Yuan 2021;Komnitsas et al. 2010). However, although such methods undoubtedly improve greatly the accuracy of boundary-extraction results, those studieswhich relied mostly on substantial field operations-have been vigorously challenged. The researchers had to carry out burdensome outdoor surveys, which consumed a lot of manpower, material resources, and time while being unable to handle the problem of temporal and spatial heterogeneity in the study area. In this case, it appears that using remotesensing techniques may be an excellent alternative.
With the development of remote-sensing technologies, many predecessors have used different types of satellites to extract farmland boundaries. Wassie et al. (2018) created a method that realized semi-automatic extraction of cadastral boundaries using WorldView-2 satellite images. Xue et al. (2021) extracted crop boundaries in GF-2 remote-sensing images based on an improved watershed segmentation algorithm. Watkins and Van Niekerk (2019) modified the objectbased image analysis (OBIA) method to realize automatic identification and delineation of agricultural fields from Sentinel-2 imagery. However, satellites have shortcomings, such as high cost, long operation period, and severe restrictions of meteorological conditions (Zheng et al. 2018). In recent years, the emerging technology of unmanned aerial vehicles (UAVs) has offered a series of advantages, such as flexibility, low cost, short data-acquisition cycle, and low risk (Hu and Li 2019). UAVs combined with satellite images allow a novel perspective on target monitoring as well as overcoming the aforementioned deficiencies of conventional remote-sensing methods to some extent (Martin et al. 2018;Lu et al. 2021). This combination has been used in many large-scale data acquisitions (Gao et al. 2021;Li et al. 2020). In small-scale studies, given that high-precision satellite images are often difficult to obtain, the flexibility and relatively high resolution of UAVs make them more economical for small-scale tasks. Therefore, using UAV technology to obtain the boundaries in areas of coal-mining subsidence would enhance the timeliness of extraction, reduce the workload of field workers, and provide technical support for the subsequent establishment of large-scale models combined with satellite data (Xiao et al. 2017;Park and Choi 2020).
Because of their high efficiency and stability, vegetation indices (VIs) are the parameters used most commonly to estimate crop growth in remote sensing. Meanwhile, studies have shown that the values of above-ground biomass (AGB), soil and plant analysis development (SPAD), and leaf area index (LAI) are strongly related to crop growth and yield, thereby meaning that they are important tools for agricultural monitoring (Villoslada Peciña et al. 2021;Zhao et al. 2021). Nowadays, the images acquired using cutting-edge hyperspectral data contain richer spatial, radiant, and spectral information and the spectral resolution has been greatly improved, but the cost is high (Chen et al. 2019). Furthermore, as the number of bands increases, so the amount of data increases exponentially, which leads to information redundancy and difficulty in post-processing. Therefore, in the present study, relatively cheap and simply processed multispectral data were used to construct the corresponding VIs for analyzing crop growth and yield.
To date, UAV remote-sensing technology based on multispectral data has been used widely in land surveys and for evaluating land damage and crop production, and UAVs have already played an important role in China's third national land survey (Li 2021). Using UAV multispectral remotesensing data, Padro et al. (2019) developed a rapid, effective, and non-destructive method for evaluating the quality of restorations after the damage done to land by open-pit mining. Jełowicki et al. (2020) used a multispectral camera to evaluate a frost-damaged area, and they used the OBIA method to remove the impact of roads. To assess the impact of climate disasters on crops and reasonably determine agricultural insurance, Vlachopoulos et al. (2020) recommended pixel-based and object-based random forest (RF) algorithms to extract boundaries using multispectral data obtained by UAVs. However, although many scholars have conducted much related research on the problems of land damage, few studies have addressed directly the extraction of boundaries with various damage levels. Threshold segmentation is commonly used in image processing (Singh et al. 2015;Xie et al. 2010;Lei and Fan 2015) the principle of this simple method being to select an appropriate threshold determines whether each pixel meets the threshold requirements and then divides all the pixels into respective intervals that meet the requirements. Another robust and popular algorithm is Otsu segmentation, which is convenient to calculate, is unaffected by image brightness and contrast, and so has been used widely in image processing ). However, the traditional Otsu algorithm calculates the gray values of images, is generally applied to discrete data, and so cannot be used directly for threshold segmentation and needs to be improved. Therefore, in the present study, an improved Otsu algorithm was introduced to extract reduced-production boundaries (RBs).
Based on the above analysis, the present study took the Dongtan coal mine in China's Shandong province as a representative CGOA with high groundwater levels and used a DJI M100 UAV equipped with a Parrot Sequoia multispectral camera and a ZenmuseX3 camera for aerial photography. Moreover, field sampling was also done during the study period. The aims were as follows: (i) to construct a digital orthophoto map (DOM) and digital surface model (DSM) of the study area, analyze the distribution of the surface elevation, and perform elevation threshold segmentation to extract the OB and (ii) to use four regression methods-power regression (PR), exponential regression (ER), multiple linear regression (MR), and RF-to estimate the values of AGB, SPAD, and LAI. After optimizing the inversion model, a reduced-production map (RM) was constructed, and finally, RBs were extracted by an improved Otsu algorithm. In summary, the present study is useful for subsequent reclamation work and provides a solid technical foundation for research on determining compensation for damaged crops.

Study area
The study was conducted at the Dongtan coal mine in Jining city (116°50′49″E-116°56′56″E, 35°24′11″N-35°31′25″N). As a typical CGOA with high groundwater levels, the  cornfield of the A1 workface was selected as the study area ( Fig. 1). This region is characterized by a warm monsoon climate with an average annual temperature of 13.3-14.1 °C and an average annual precipitation of 597-820 mm. The research area slopes gently downward from northeast to southwest with the ground elevation between + 42.66 and + 54.58 m. The water table is quite shallow and fluctuates between 2 and 3 m all year round.
The mining of the working face started in August 2014. During the field study in August 2017, the average mining depth was 550 m and the maximum subsidence depth reached 7 m, so the surface settlement in most of the affected areas exceeded the average phreatic level. In this case, a body of groundwater flowed up and flooded a large amount of high-quality cultivated lands, forming a unique waterlogged landscape. Furthermore, the surface of subsidence within the study area was stabilized during field campaigns. Meanwhile, to complete the field survey and sampling work, it was ensured that this study area had not been repaired by manual restoration beforehand.

UAV remote-sensing data collection
The data in the study were acquired by a DJI Matrice100 UAV (DJI Technology Co., Shenzhen, China) equipped with a Parrot Sequoia multispectral camera (Parrot, Switzerland). The four bands of this camera are green (530-570 nm), red (640-680 nm), red edge (730-740 nm), and near infrared (770-810 nm), each with 1.2 million pixels (1280 × 960). In addition, the UAV was also equipped with a ZenmuseX3 digital camera (DJI Technology Co., Shenzhen, China) with 16 million pixels, which realized the simultaneous collection of visible-light data and terrestrial spectrum information. The flight was performed on 12 August 2017 at an altitude of 110 m above the ground, a speed of 9 m/s, and with a spatial resolution of 13 cm. The area covered during the flight mission was ~1.1 km 2 , and a total of 4980 photos were collected, completely covering the entire study area.
The images were processed in the Pix4DMapper software (Pix4D SA, Switzerland). One excellent feature of this software is that the image stitching and geometric and radiometric corrections are all completed automatically, while the DOM and DSM can be acquired quickly without extra human knowledge.

Field data collection
Field sampling was carried out on 13 August 2017. Based on mining subsidence theory and the probability integration method, combined with various surface-movement parameters (sink coefficient, horizontal movement coefficient, etc.) and coal-seam information (coal-seam thickness, inclination, etc.), the mining subsidence prediction software (MSPS) was used to predict mining subsidence. This software is widely used in the prediction of coal mining subsidence and damage in China. Meanwhile, according to the traditional definition, the 10-mm subsidence contour was used as the basis of disturbed boundaries. Combining the two approaches above, the subsidence contour of the working face corresponding to the study area was achieved. Three sampling lines-i.e., L1 (550 m), L2 (540 m), and L3 (620 m)-were laid out along the strike, dip, and angular bisector directions. The spacing of the sample points was 5 m, 10 m, 20 m, 30 m, and 60 m in sequence, generating a total of 54 sampling points (Fig. 1). For each sampling point, a 1 m × 1 m standard plot was designed to collect AGB, SPAD, and LAI values.

Measurement of AGB
The AGB was determined by "the field harvest method" (Liu et al. 2020). Two maizes (Maize 1 and Maize 2) were selected randomly in the plot as the test samples. In the laboratory, the two samples were stored for 10 h at 105 °C and dried at 80 °C for 24 h until reaching their constant weights. The final AGB was calculated using where N is the number of crop in each sample plot, S [m 2 ] is the area of the sample plot, and AGB 1 and AGB 2 are the maize AGB values of samples 1 and 2, respectively.

Measurement of chlorophyll contents
This parameter was measured with SPAD-502 using the "fivepoint method." Five uniformly growing maizes were selected, and the leaves along the tip, middle, and base of each plant were measured from the top leaf to the secondary leaf twice. Finally, the measured values of the five plants in the plot were averaged and used as the SPAD value for the sample point.

Measurement of LAI
To determine the LAI, the most traditional "length-width coefficient method" was selected (Kross et al. 2015). The length and maximum width of all the leaves of five plants were measured using a steel ruler, and the leaf-area correction coefficient for the north China plain, i.e., 0.7017, was used. The formula is where k is the leaf-area correction coefficient, L is the leaf length, B is the leaf width, n is the total number of leaves of plant j , m is the number of selected plants, and is the planting density.

Elevation threshold segmentation
Flooded crops and unaffected crops show significant elevation differences, based on which the desired OB could be achieved. Threshold segmentation segments the background and the target by determining a specific threshold; it is a simple segmentation algorithm (Kaur and Kaur 2017). The threshold segmentation function is defined as where T is the predetermined threshold, the pixel of the set target is g(i, j) = 1 , and the others are g(i, j) = 0 . Thus, pixels with larger values of the given feature belong to the foreground group, while those with smaller values are placed in the background group.
For the OB identification task, ArcGIS 10.2 was used to make multiple profile lines on the DSM of the study area, and then discontinuity points of each profile line were recorded, using the average value of these points as the threshold for elevation segmentation. The data with elevation higher than the threshold were treated as the non-outof-production area, while the data below the threshold were classified as the out-of-production area.

Optimization of vegetation index
There is a strong correlation between the characteristic bands of spectral reflectance and vegetation photosynthetic pigments. Therefore, VIs based on this relationship are good tools for testing the health of plants and solving the problem of defining the scope of damaged cultivated land (Jełowicki et al. 2020;Kross et al. 2015). We selected some VIs that have been commonly used in previous research and that met the requirements of the present study, and on this basis, considering the unique performance of the red-edge band of the spectrum (Zhao et al. 2020a, b), we replaced the red and green bands with this special spectral trait to improve and expand the existing VIs during the study. Through correlation analysis of AGB and VIs, the rededge microwave VI (MVI redge ), the normalized difference VI (NDVI), and the red-edge modified simple ratio index (MSR redge ) were selected as the optimal VIs for biomass estimation. For the study of SPAD, the red-edge NDVI (NDVI redge ), the green NDVI (GNDVI), and the red-edge chlorophyll index (CI redge ) were chosen. Meanwhile, the rededge-modified simple ratio index (MSR redge ), the GNDVI, and the green chlorophyll index (CI green ) were used to estimate the LAI (Table 1).

Modeling and evaluation
In this study, four popular regression models-i.e., PR, ER, MR, and RF-were used for the AGB, SPAD, and LAI analyses (Fig. 2). Aiming to avoid the negative influence of coarse errors in collected field data, triple standard difference method (Huang et al. 2006) was conducted to refine the data, whereupon these datasets without gross errors were subjected to random sampling. In this stage, the data were divided into two different groups, i.e., a modeling set and a validation set, to evaluate the modeling accuracy from the perspective of cross-validation (Mahmoodzadeh et al. 2020;Maimaitijiang et al. 2019). Subsequently, the rootmean-square error (RMSE) and the coefficient of determination (R 2 ) were selected to support the verification of model accuracy. R 2 represents the degree of fitting between the simulated and measured values; the closer R 2 is to 1, the better the degree of fitting. The RMSE reflects the degree of deviation between the simulated and measured values; the smaller the RMSE, the more accurate the representation.
In this study, the final selected optimal inversion models of AGB, SPAD, and LAI combined with the DSM were used to perform equal-weight spatial calculations. As a result, the RM was generated by overlaying the characteristics of these four sources. The value of the RM, i.e., the reducedproduction coefficient, was an imperative criterion for gauging crop growth: the lower the coefficient, the more serious the reduction in corn yield.  Gitelson et al. (1996)

Otsu threshold segmentation
The Otsu algorithm was proposed by Otsu (1979); it is an efficient algorithm for image binarization and can maximize the between-class variance between foreground and background images. Of all segmentation methods, the Otsu method is one of the most successful for image thresholding because of its simple calculation (Vala and Baxi 2013). The basic idea is as follows. The total number of pixels in the given image is N with gray levels [0, L − 1] . There are ni pixels corresponding to gray level i , and its probability can be expressed as There is a threshold T , and the gray values are divided into two categories C 0 and C 1 according to T . The average value of the distribution probability based on the gray levels of the entire image is expressed as The average values C 0 and C 1 are where From (5)- (7), it can be deducted that And it can be concluded that the between-class variance of the image is Combining (7)-(9) gives And combining (9) and (10) gives T is selected sequentially in the range [0, L − 1] , and the value of T that maximizes the between-class variance 2 B is the optimal threshold for the Otsu algorithm.

Fig. 2 Scheme of present study
However, the images of reduced production in the study area comprised raster data and double-type discrete data, in which case the traditional Otsu algorithm could not be used directly for threshold segmentation and had to be improved as follows.
1. Denote the minimum and maximum values of the given image as MinData and MaxData, respectively, select the initial segmentation number breaknum 0 , and the corresponding unit interval Step 0 is obtained as 2. Based on the traditional Otsu algorithm, take the value between MinData and MaxData using Step 0 , find the maximum inter-cluster variance 0 under breaknum 0 and the corresponding value of the reduced-production degree coefficient h0 , and select a new NNi = 2 × N0 at the same time.

Recalculate (12) to generate a new i and corresponding
hi . When hi − hi − 1 is less than the threshold, the iteration stops.

OB extraction
The values of the discontinuity points of the four selected profile lines a-d (Fig. 4a) were 46.41 m, 46.52 m, 46.98 m, and 46.01 m, respectively (Fig. 3), and the calculated average of 46.48 m was taken as the threshold for the elevation segmentation. Then, after some optimized operations, mathematical morphology, boundary tracking, and manual adjustments, the OB was extracted successfully (Fig. 4f).

Estimation of AGB, SPAD, and LAI
Two common and popular regression algorithms-i.e., PR and ER-were used to find the optimal simple regression VI according to the value of R 2 (Fig. 5). In addition, combining MR and RF, the best inversion model for AGB, SPAD, and LAI was further explored (Fig. 6). The number of processed samples for AGB, SPAD, and LAI was 36, 35, and 36, respectively, and each plant parameter selected 10 samples into validation dataset (Table 2) by using hold-out cross-validation method (HOCV). In the experiments for the inversion modeling operation of AGB and LAI, we found that the modeling accuracy of RF was the highest (Table 2). In the validation set, their RMSEs showed signs of decline, but the R 2 of AGB was not promising. To eliminate this type of uncertainty, we then used leave-one-out cross-validation (LOOCV) to examine all the inversion models. The results showed that the predicted and measured values of AGB and SPAD using RF were in good agreement (Table 2), indicating that the RF method significantly improves the inversion accuracy of agricultural AGB and SPAD. However, in the modeling process of LAI, the accuracy of the MR model was higher than that of the RF model, and R 2 and RMSE were 0.88 and 1.07, respectively.
Thus, through the above analysis, we concluded that RF is more effective than the other competitors for estimating AGB (using MVI redge , NDVI, MSR redge with R 2 = 0.83, RMSE = 0.114 kg·m −2 ) and SPAD (using NDVI redge , GNDVI, CI redge with R 2 = 0.83, RMSE = 0.152 SPAD), which showed higher R 2 and lower RMSE during the test. In addition, MR (using MSR redge , GNDVI, CI green with R 2 = 0.88, RMSE = 1.070) exhibited stronger agreement for LAI estimation, although its RMSE was higher than that of RF, and the lower R 2 emphasized its exemplary performance. The comparison of best single VI estimation and combined VI estimation is presented in Table 2 and Fig. 6.   Fig. 6 Relationship between UAV-derived and field-measured parameters

RB extraction
The estimation results showed that the values of AGB, SPAD, LAI in the study ranged between 10.05 ~ 1847.42 g·m −2 , 6.25 ~ 63.57 SPAD, and 0.15 ~ 14.41 respectively (Fig. 7), which were consistent with the measured results collected from field samples. The RM showed that the reduced-production coefficient of corn fluctuated between 0.18 and 2.79. In more detail, within the area covered by L1, this coefficient was 0.6-2.2, and the average value was below 2. L2 and L3 covered all levels of the reduced-production coefficient, and at the farthest ends of sample lines L2 and L3, the value of the reduced-production coefficient reached 2.3-2.79. Compared with the modeling results of a single parameter, the spatial difference in the distribution of multiple parameters was more significant (Fig. 7). Clearly, the coefficient values were increasing gradually from the center of the subsidence basin toward the outside. In addition, Fig. 7 shows that cultivated land outside the disturbance boundary was also slightly disturbed. However, according to the field investigation, this was mainly because of the regional cultivation conditions and different crop-care approaches in the study area, and this effect does not interfere with the experimental results.
Subsequently, the improved Otsu algorithm was used to perform threshold segmentations on the RM, and the severe RB and mild RB in the study area were obtained only after performing mathematical morphology and boundary-tracking methods. The two different levels of RBs and OB were combined in the DOM, and the full results of OB and RB extraction are shown in Fig. 8e.

Performance of elevation threshold segmentation for OB
The study area is a typical CGOA with high groundwater levels, and the OB is located between the flooded tidal flat and the unflooded land. In this case, the flooded crops and the unaffected crops show significant elevation differences in the DSM. As a result, using this special characteristic based on the simple segmentation algorithm, the OB can be extracted accurately. In the follow-up exploration, aiming to verify the effect of boundary extraction, we selected four small research areas A-D to probe for more details (Fig. 9). The outer edge of the red line is obviously darker than other cultivated land. In this scenario, the dark area represents the direct result of abnormal crop growth, and this background gap provides a convenient and intuitive way to verify that the extracted boundary coincides well with the actual out-of-production edge. The Otsu segmentation algorithm can also be used to extract the OB, and Zhao et al. (2020a, b) provided a valid reference by implementing Otsu segmentation based on RGB images. However, this trait on the elevation is obvious, and good results can be achieved by simple threshold segmentation; the most important component of the extraction results should be the accuracy of the DSM. In addition to using RGB images, this study also added multispectral bands to construct the DSM, so the accuracy of the extraction results was greatly improved.

Performance comparison of models for AGB, SPAD, and LAI estimation
This study used multispectral and visible-light images collected by a UAV, combined with field-measured data, to extract boundaries of disturbed cultivated land in different levels. Three important parameters that are closely related to vegetation growth (AGB, SPAD, and LAI) were selected to establish an RM before final extracting the results. Commonly used regression models (PR, ER, MR, and RF) were used during this step, and the outperformance of RF in the estimation of AGB and SPAD compared to MR and PR (Table 2) is supported by many previous studies. Liang et al. (2019) used MR, a support vector machine, an artificial neural network, and RF to create a suitable model for AGB estimation, and the RF model gave the most balanced results, with low error and a high ratio of explained variance for both the training and test sets. Feng et al. (2018) used RF, partial least-squares, a backpropagation neural network, and a support vector machine to estimate the chlorophyll content of apple leaves, and RF was regarded as the best modeling choice. In other work, MR showed promising results as an estimator for LAI as supported by studies of LAI estimation of a natural forest (Pu 2012) and one-winter wheat growth (Han et al. 2021).

Performance of improved Otsu segmentation algorithm for RBs
In the RM, the spatial distribution of coefficient showed a certain law (Fig. 7d), i.e., the closer to the out-of-production area, the lower the value. Hence, its coefficient increased gradually from the inner area of the subsidence basin to the outside along the direction of the three sampling lines. Based on this, RBs could be obtained by applying the improved Otsu segmentation twice. Although many scholars have invented miscellaneous improved Otsu methods to date (Sun et al. 2016;Rajinikanth and Couceiro 2015), there is little trace of them being used to extract the boundaries of cultivated land.
Moreover, to analyze the accuracy of the RBs, in Fig. 9, we magnified area E to identify the performance of the RBs, but the comparison between the two levels of RB was not satisfactory. In this study, although we reason that the actual yield of crops is the most reasonable and effective test standard, weather conditions deprive this method of the necessary objectivity. Because after completion of the field sampling, Shandong province was affected in the following week by the severe typhoon Tiange (which started on August 24, 2017), making any ground-based attempt to gauge the yield impossible. Hence, using actual yield data as a standard to test the accuracy of boundary extraction should be investigated further.

Limitations and future research directions
Despite ideal results having been shown, the following limitations require further consideration. During the dataacquisition period, only one flight was conducted. Also, the crop-growth parameters and the UAV flight data were separated by a day, so the fitness of these datasets may have been disturbed. In addition, we selected a summer CGOA as the research object, this being because many uncertain variables throughout the year can interfere with crop growth, and it is difficult to avoid the influence of season on the results. Moreover, the VIs and modeling functions that we chose are relatively limited. Also, in the process of boundary extraction, the road was relatively absent of vegetation, which could have led to a relatively low reduced-production coefficient, thereby possibly compromising the accuracy of the directly extracted boundary.
In terms of the accuracy of boundary extraction, because of the typhoon in the subsequent stage, we did not collect the actual output of crops as our most powerful evaluation standard. Therefore, our follow-up research will develop and establish a complete data-acquisition plan and system to eliminate the deficiencies in acquisition time and boundary verification and continue to optimize inversion models from the perspectives of data sources, VIs, and modeling methods. At the same time, we will explore how to apply this research idea to achieve fast and precise boundary extraction at larger scale. UAVs combined with satellites will become a focus of future study.

Conclusion
This paper presents an efficient approach for using lowcost, time-saving, high-resolution RGB and multispectral images obtained from a UAV to accurately extract disturbed boundaries in a CGOA with high groundwater levels. The demonstrated methodology includes analysis of elevation to exert a simple threshold segmentation, selecting optimal models for AGB, SPAD, and LAI estimations, and approaches for using an improved Otsu method, which can be perceived as an effective alternative for extracting disturbed boundaries during the refurbishment and reclamation process.
The non-crop area and the crop area within the scope of subsidence were obviously different in elevation. The elevation threshold could be set to segment the DSM to achieve the purpose of extracting the OB accurately in the study area.
Furthermore, during the process of establishing the RM, three parameters (AGB, SPAD, LAI) highly related to crop growth were used. The model of RF with the indices MVI redge , NDVI, and MSR redge as variables proved to be the best estimator for AGB. The model of RF with the indices NDVI redge , GNDVI, and CI redge outperformed the other combinations for SPAD estimation. In the inversion test of LAI, the MR model with the indices MSR redge , GNDVI, and CI green tended to be a superior choice.
Based on the RM, two different RBs were extracted successfully by implementing an improved Otsu segmentation method. This study offers an innovative insight into using simple physiological traits of crops to identify their disturbed boundaries swiftly and accurately, as well as relieving the heavy workload during the reclamation and compensation survey.