Sugarcane Phenology Retrieval in Heterogeneous Agricultural Landscapes Based on Spatiotemporal Fusion Remote Sensing Data

Abstract

Accurate phenological information on sugarcane is crucial for guiding precise cultivation management and enhancing sugar production. Remote sensing offers an efficient approach for large-scale phenology retrieval, but most studies have primarily focused on staple crops. The methods for retrieving the sugarcane phenology—the germination, tillering, elongation, and maturity stages—remain underexplored. This study addresses the challenge of accurately monitoring the sugarcane phenology in complex terrains by proposing an optimized strategy integrating spatiotemporal fusion data. Ground-based validation showed that the change detection method based on the Double-Logistic curve significantly outperformed the threshold-based approach, with the highest accuracy for the elongation and maturity stages achieved at the maximum slope points of the ascending and descending phases, respectively. For the germination and tillering stages with low canopy cover, a novel time-windowed change detection method was introduced, using the first local maximum of the third derivative curve (denoted as Point A) to establish a temporal buffer. The optimal retrieval models were identified as 25 days before and 20 days after Point A for germination and tillering, respectively. Among the six commonly used vegetation indices, the NDVI (normalized difference vegetation index) performed the best across all the phenological stages. Spatiotemporal fusion using the ESTARFM (Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model) significantly improved the monitoring accuracy in heterogeneous agricultural landscapes, reducing the RMSE (root-mean-squared error) by 21–46%, with retrieval errors decreasing from 18.25 to 12.97 days for germination, from 8.19 to 4.41 days for tillering, from 19.17 to 10.78 days for elongation, and from 19.02 to 15.04 days for maturity, highlighting its superior accuracy. The findings provide a reliable technical solution for precision sugarcane management in heterogeneous landscapes.

1. Introduction

Sugarcane, characterized by its high sucrose concentration in the stem internodes, stands as the world’s predominant source of sugar, accounting for over 70% of the global production [1,2,3]. Cultivated extensively in tropical and subtropical regions with demands for abundant heat and water, it is also increasingly vital as a bioenergy feedstock [4,5]. Understanding the sugarcane phenology is essential not only for optimizing critical management practices (e.g., fertilization, irrigation, pest control, harvesting) but also for accurate crop growth modeling and yield forecasting. Furthermore, the sugarcane phenology serves as a crucial bio-indicator, reflecting the plant’s physiological responses to environmental changes. This provides critical insights into how climate change impacts sugarcane growth and productivity [6,7].

Compared to the traditional field observations, which are both time-consuming and labor-intensive, remote sensing offers an efficient approach for the large-scale monitoring of crop phenology. Phenological information can be acquired by analyzing the dynamics of spectral signals [8]. Retrieving crop phenology from remote sensing data typically involves time series reconstruction and the extraction of mathematical metrics [9]. To mitigate noise caused by cloud contamination, shadow effects, and other radiometric errors, methods such as the moving average filter [10], Double-Logistic function fitting [11], and HANTS [12] are generally applied to generate more reliable time series data. Phenology information is then commonly derived from the reconstructed time series using threshold-based methods [13,14,15] and change detection methods [11,16]. The threshold-based methods determine phenological dates by applying predefined relative or absolute thresholds [17], while the change detection methods identify these dates by detecting inflection or transition points along the fitted curve [11]. These approaches provide robust tools for analyzing phenological patterns on a regional or global scale. In addition to phenology retrieval strategies, the selection of suitable remote sensing indices also plays a critical role [18]. The commonly used vegetation indices include spectral indices such as the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), enhanced vegetation index 2 (EVI2), soil-adjusted vegetation index (SAVI), ratio vegetation index (RVI), and green–red vegetation index (GRVI) [19,20]. Physical parameters, such as the fraction of absorbed photosynthetically active radiation (FPAR) [21,22] and leaf area index (LAI) [23,24,25], are also widely utilized.

While crop phenology estimation has been extensively studied for staple crops such as paddy rice, maize, soybean, and wheat, methods for sugarcane remain underdeveloped. Sugarcane’s life cycle comprises four distinct phenological stages: germination (initial shoot emergence with sparse canopy), tillering (the emergence of multiple stalks from the primary shoot), elongation (rapid stalk extension and biomass accumulation), and maturity (photosynthetic decline, sucrose accumulation, and harvest readiness), which are illustrated in Figure 1. Although studies like that of Wang et al. [26] have derived remote sensing phenological parameters (e.g., green-up date, start date of peak season, senescent date), a significant gap persists between these features and actual physiological processes [18]. Crucially, the relationship between these remotely sensed parameters and the actual ground-observed phenology stages has yet to be examined, with no established metrics reliably representing sugarcane’s germination, tillering, elongation, or maturity.

As the world’s third-largest sugarcane producer, China’s sugarcane cultivation is concentrated in its southern provinces (Guangxi, Yunnan, Guangdong, and Hainan), where the complex terrain generates fragmented landscapes with high heterogeneity [26]. This topographic constraint necessitates high spatiotemporal-resolution time series remote sensing for precise phenology monitoring. Furthermore, frequent cloud cover and high rainfall in major sugarcane regions hinder optical remote sensing acquisition, thereby affecting the phenology retrieval accuracy. In previous studies, MODIS, AVHRR, and VIIRS with frequent revisit capabilities have been widely used in global and national phenology. For example, global land cover and land surface phenology products were generated using 500 m resolution MODIS data [11,27]. Sakamoto et al. [28] developed a shape model to fit the time series of the wide dynamic range vegetation index derived from MODIS data, successfully estimating the phenology of corn in the United States [29]. Luo et al. [30] created a phenology dataset for paddy rice, wheat, and maize in China from 2000 to 2015 based on the GLASS (Global Land Surface Satellite) LAI product, which was retrieved from reflectance data acquired by MODIS and AVHRR sensors. However, their coarse spatial resolution fundamentally limits its applicability in heterogeneous agricultural regions. Spectral mixture effects inherent to such resolutions significantly degrade the phenology extraction accuracy. Although medium-resolution data like Landsat, Sentinel-2, and HJ-1 A/B data provide finer spatial detail [31], narrow swaths and extended revisit intervals cause critical data gaps, particularly in persistently cloudy regions like South China, where observations may be absent for >30 days. To overcome these challenges, various spatiotemporal image fusion methods have been developed to generate time series data with high spatial resolution and frequent observation intervals. One such method is the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), an advanced version of the STARFM [32]. The ESTARFM combines pairs of Landsat and MODIS images to predict the Landsat-scale reflectance on dates when only MODIS data is available. It has demonstrated an improved prediction accuracy in heterogeneous landscapes [33]. The ESTARFM has two main mechanistic advantages: spatial decomposition, which isolates signals from mixed pixels in smallholdings, and temporal interpolation, which reconstructs cloud-free time series to capture rapid phenophase transitions. This approach holds great promise for enhancing phenology estimates of sugarcane in South China, a region characterized by complex terrain and fragmented agricultural landscapes, but its effectiveness still needs to be validated through experimental studies.

To address these unresolved issues, this study aimed to (1) determine the most effective retrieval strategies for the key sugarcane phenological stages by comparatively evaluating the threshold-based and change detection methods against ground-observed phenological records, (2) evaluate the performance of multiple vegetation indices for retrieving each sugarcane phenological stage and determine the most suitable vegetation index for sugarcane phenology estimation, and (3) quantitatively evaluate whether ESTARFM-derived spatiotemporal fusion data enhances the sugarcane phenology estimation accuracy in heterogeneous landscapes, relative to conventional low-resolution time series data.

2. Materials

2.1. CMA Observation Data

Sugarcane phenology observation records were obtained from the agricultural meteorological stations operated by the China Meteorological Administration (CMA). The phenological data were collected in experimental fields managed by the CMA. A total of seven sugarcane observation stations were selected across the primary sugarcane-producing regions in South China (Figure 2). Five stations were located in the Guangxi Zhuang Autonomous Region, the largest sugarcane-producing area in China: Fusui Station, Guigang Station, Shatang Station, Yizhou Station, and Pingguo Station. Two stations were located in Yunnan Province, the second largest sugarcane-producing area in China: Gengma Station and Yuanjiang Station. These CMA stations conducted ground-based observations of the sugarcane phenology, recording the key phenological dates for germination, tillering, elongation, and maturity from 2001 to 2012.

2.2. Multisource Satellite Data

Two satellite data sources were utilized in this study, including the daily MODIS reflectance product and 30 m resolution Landsat data. Specifically, the daily MODIS NBAR (Nadir BRDF-Adjusted Reflectance) product (MCD43A4) includes 500 m reflectance data adjusted for nadir viewing using a bidirectional reflectance distribution function (BRDF). The product integrates data from both Terra and Aqua satellites, generating daily observations by selecting the most representative pixel within a 16-day period. Low-quality pixels were excluded using the product’s quality band. From the MODIS multispectral bands, the blue, green, red, and near-infrared bands were chosen. These MODIS reflectance data were bilinearly resampled to match the 30 m resolution of Landsat. Landsat 7 Level 2 surface reflectance data with total cloud cover less than 70% were initially selected, and the cloud-contaminated pixels were then masked using the quality band to retain the valid pixels, which could sufficiently utilize fragmented valid observations through cloud gaps in cloudy and rainy regions. Both the MODIS and Landsat products were acquired from the Google Earth Engine (GEE) platform (https://earthengine.google.com/, accessed on 25 January 2024).

3. Methods

The methodological framework of this study for sugarcane phenology retrieval comprises five main components: (1) ESTARFM spatiotemporal fusion using Landsat–MODIS image pairs, (2) time series reconstruction for multiple vegetation indices, (3) Double-Logistic function fitting, (4) phenological metrics extraction using the change detection method and threshold-based method, and (5) accuracy assessment based on ground-observed phenological dates (Figure 3).

3.1. ESTARFM Fusion for Generating High-Spatiotemporal-Resolution Time Series Data

The ESTARFM spatiotemporal fusion method was applied to fuse MODIS and Landsat data, producing daily Landsat-like time series data. The ESTARFM assumes that the reflectance of mixed coarse-resolution pixels is a linear combination of the reflectance of multiple land cover types. The weights for each land cover type are determined by their fractional area coverages, and it is assumed that the proportions of land cover types within the mixed pixels remain constant during the prediction period. The algorithm first searches for similar pixels within the same land cover type nearby. It then weights these pixels based on their spatial and spectral proximity. Finally, conversion coefficients are computed to predict the fine-resolution reflectance when only coarse-resolution data exists [33]. The Landsat data were paired with MODIS data based on the nearest observation dates. Cloud-free images were prioritized in the selection of the image pairs.

3.2. Reconstruction of Multiple Vegetation Index Time Series

In this study, multiple vegetation indices were calculated based on the fused dataset, including the NDVI (normalized difference vegetation index) [34], EVI (enhanced vegetation index) [35], EVI2 (enhanced vegetation index 2) [36], SAVI (soil-adjusted vegetation index) [37], RVI (ratio vegetation index) [34], and GRVI (green–red vegetation index) [34]. The NDVI, derived from the near-infrared band and red bands, is one of the most widely used indices for monitoring the vegetation status. It provides a comprehensive indication of the greenness and vegetation fraction cover, reflecting both the chlorophyll content and canopy structure [34]. The EVI is a more complex index that incorporates the near-infrared, red, and blue bands [35]. It has the advantage of reducing atmospheric influence and minimizing soil background interference and is more sensitive in areas with dense vegetation or high biomass, where the NDVI tends to saturate. The EVI2, a simplified version of the EVI, is designed for cases where the blue band is unavailable, while still retaining the benefits of reduced atmospheric and soil background influences. The EVI2 has been used in generating VIIRS land surface phenology (LSP) products [36]. The SAVI includes a soil adjustment factor (L), which makes it more reliable in areas with sparse vegetation, where the soil influence is significant [37]. The RVI is calculated as the ratio of the near-infrared to red band, multiplied by a constant factor (α) to adjust the value range. The GRVI is particularly sensitive to chlorophyll content, as it measures the contrast between the green and red bands [34]. Vigorous and healthy vegetation typically has higher chlorophyll concentrations, which result in increased green-light reflection and red-light absorption. The equations of each vegetation index are shown in Equations (1)–(6):

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\rho }_{nir}-{\rho }_{red}}{{\rho }_{nir}+{\rho }_{red}}}$$

(1)

$$\mathrm{E}\mathrm{V}\mathrm{I}={\displaystyle \frac{2.5\times ({\rho }_{nir}-{\rho }_{red})}{{\rho }_{nir}+6\times {\rho }_{red}-7.5\times {\rho }_{blue}+1}}$$

(2)

$$\mathrm{E}\mathrm{V}\mathrm{I}2={\displaystyle \frac{2.5\times ({\rho }_{nir}-{\rho }_{red})}{1+{\rho }_{nir}+2.4\times {\rho }_{red}}}$$

(3)

$$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}={\displaystyle \frac{(1+L)\times ({\rho }_{nir}-{\rho }_{red})}{{\rho }_{nir}+{\rho }_{red}+L}}$$

(4)

$$\mathrm{R}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\rho }_{nir}}{{\rho }_{red}}}\times \alpha$$

(5)

$$\mathrm{G}\mathrm{R}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\rho }_{green}-{\rho }_{red}}{{\rho }_{green}+{\rho }_{red}}}$$

(6)

where ${\rho }_{\mathit{blue}}$, ${\rho }_{\mathit{green}}$, ${\rho }_{\mathit{red}}$, and ${\rho }_{\mathit{nir}}$ represent the reflectance of the blue, green, red, and near-infrared spectral bands, corresponding to Bands 1, 2, 3, and 4 of Landsat 7, respectively. L (L = 0.5) and α (α = 0.1) are the adjustment factors in the equations.

The maximum value composite approach was applied to the time series of each vegetation index, producing a more reliable time series with 8-day intervals. To further reduce noise in the time series, outliers were filtered out and missing values were interpolated. Specifically, an 11-width sliding window was constructed, and if the central observed value was 0.2 lower than the median value within the sliding window, it was identified as an outlier and discarded. After that, the spline function was used to interpolate the missing values along the time series. As a result, high-quality vegetation index time series data with an 8-day observation frequency was obtained.

3.3. Double-Logistic Function Fitting for Vegetation Index Time Series

To model the seasonal dynamics of sugarcane, the Double-Logistic function was applied to fit the vegetation index time series. This function consists of an ascending phase and a descending phase, corresponding to the growth and senescence stages of the vegetation life cycle, respectively (Figure 3) [7,11,17]. The ascending phase can be divided into the onset of green-up, rapid growth, and peak growth stages. During the green-up onset, vegetation begins to grow, and photosynthesis activity increases. In the rapid growth phase, the chlorophyll content increases, canopy expansion occurs, and the photosynthesis rate accelerates, while the peak growth stage marks the maximum greenness. The descending phase, which represents the decline in vegetation growth, is also divided into three stages: the onset of senescence, rapid senescence, and late senescence stages. During the onset of senescence, physiological activity slows, with a decrease in the chlorophyll content and canopy cover. The rapid senescence stage is presented as accelerated aging, and the late senescence stage reflects a significant reduction in the vegetation index, indicating that the senescence process is largely complete.

The expression of the Double-Logistic function is shown in Equation (7):

$$V(t)=V_{\min a}+{\displaystyle \frac{V_{\max }-V_{\min a}}{1+e^{-S_a\cdot (t-T_a)}}}+{\displaystyle \frac{V_{\max }-V_{\min b}}{1+e^{-S_b\cdot (t-T_b)}}}$$

(7)

where V(t) represents the value of the modeled curve at time (t). V_max is the maximum value during the vegetation growth cycle. V_mina and V_minb represent the minimum values of the curve in the ascending and descending stages, respectively. V_max − V_mina and V_max − V_minb are the amplitudes of the ascending and descending parts, respectively. T_a and T_b are the times with the maximum slope during the green-up and senescence stages, respectively, while S_a and S_b are the slopes at T_a and T_b, respectively. In the solution of these unknown parameters, the values of V_mina, V_minb, and V_max are first determined from the baseline and peak values of the time series. The logistic equations for the ascending and descending phases are then separately transformed into linear equations, and the unknown parameters are estimated using the least-squares method.

3.4. Phenological Metrics Extraction from Vegetation Index Time Series

3.4.1. Change Detection Method

The change detection method identifies phenological dates by capturing the dynamic characteristics of the fitted time series (Figure 4). The first derivative of the time series was calculated to capture the point of the maximum increase rate (Point B) and the point of the maximum decrease rate (Point E), which represent the most vigorous growth stage and accelerated aging stage, respectively. Additionally, the rate of change in the curvature (third derivative) was computed to determine the start of the green-up phase (Point A) and the end of the peak growth phase (Point C), as well as the start of the senescence phase (Point D) and the end of the accelerating senescence phase (Point F) [11].

To explore the relationship between these key mathematical points and the sugarcane phenology observed on the ground, we tested various strategies and compared their accuracies. Sparse sugarcane canopies in the early growth stages (germination and tillering) cause strong soil background noise in spectral data. Consequently, the remote detection of phenological development becomes challenging. Therefore, temporal buffers with empirically determined widths were applied, referred to as the time-windowed change detection method in this study. Specifically, time intervals of 20, 25, and 30 days prior to Point A were applied to retrieve the germination stage. Similarly, 20, 25, and 30 days after Point A were used to retrieve the tillering stage. The elongation phase, characterized by rapid growth and the highest rate of photosynthesis activity, was directly identified by Point B with the maximum growing rate. The maturity stage, characterized by accelerated aging with decreased leaf coverage and brown leaf, was directly identified by Point E.

3.4.2. Threshold-Based Method

The threshold-based method identifies key phenological metrics by setting either a fixed or dynamic threshold within the smoothed time series. This method assumes that a phenological stage begins when the vegetation index value reaches a certain value. Since the dynamic threshold approach is more flexible in accounting for variations in the soil background across different regions, this study utilized relative ratios to retrieve the four phenological stages of sugarcane. The commonly used relative ratios of 5%, 10%, 20%, 30%, and 50% were applied in this study (Figure 5). Based on the time series fitted by the Double-Logistic function, the germination date was determined by testing the ratios of 5% and 10% of the amplitude, the tillering date by testing ratios of 20% and 30%, and the elongation date by testing the ratio of 50% during the green-up phases. The maturity dates were identified using the ratio of 50% of the amplitude during the senescence phases.

This study utilized the commonly used vegetation index, the NDVI, as an example to compare the performances of the change detection method and the threshold-based method for retrieving four phenological stages of sugarcane.

3.5. Accuracy Assessment

The accuracy of the satellite-derived sugarcane phenological dates was evaluated based on the phenological dates observed by the CMA stations. The bias and root-mean-squared error (RMSE) were calculated using the formulas in Equations (8) and (9). The bias indicates the systematic deviation between the estimated and true values, while the RMSE measures the average magnitude of the deviations between the estimated and true values:

$$\mathrm{Bias}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n({\widehat{y}}_i-y_i)}$$

(8)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{({\widehat{y}}_i-y_i)}^2}}$$

(9)

where $y_{\mathrm{i}}$ and $\widehat{y_{\mathrm{i}}}$ represent the observed phenological dates collected at the CMA stations and the estimated phenological dates for the ith record, respectively, and n is the number of valid records.

4. Results

4.1. Comparison of Different Strategies for Sugarcane Phenology Retrieval

The accuracy validation results based on the ground-observed phenology are shown in

Table 1. Accuracy of phenology retrieved from different strategies.

Phenological Stage	Retrieval Strategy	Category of Method	RMSE (Bias)/Days
Germination	Point A	Change detection	30.58 (24.31)
	20 days before Point A	Time-windowed change detection	14.62 (7.04)
	25 days before Point A	Time-windowed change detection	12.97 (2.04)
	30 days before Point A	Time-windowed change detection	13.15 (−2.96)
	5% of the amplitude	Threshold-based	28.00 (12.70)
	10% of the amplitude	Threshold-based	31.58 (22)
Tillering	Point A	Change detection	22.25 (−21.89)
	20 days after Point A	Time-windowed change detection	4.41 (−1.89)
	25 days after Point A	Time-windowed change detection	5.06 (3.11)
	30 days after Point A	Time-windowed change detection	9.04 (8.11)
	20% of the amplitude	Threshold-based	15.60 (−12.78)
	30% of the amplitude	Threshold-based	10.62 (−2.78)
Elongation	Point B	Change detection	10.78 (−3.16)
	50% of the amplitude (ascending)	Threshold-based	23.47 (8.84)
Maturity	Point E	Change detection	15.04 (7.41)
	50% of the amplitude (descending)	Threshold-based	18.50 (8.48)

and Figure 6.

For the germination stage, the results showed that the time-windowed change detection method significantly outperformed the threshold-based method. Although “Point A” presented a low performance for retrieving germination dates, with an RMSE of 30.58 days and a bias of 24.31 days, representing a systematic overestimation, the retrieval performances were obviously enhanced by tracing back a certain number of days from “Point A”. The strategies of “20 days before Point A”, “25 days before Point A”, and “30 days before Point A” all achieved better accuracies. Among them, “25 days before Point A” proved to be the most reliable method for retrieving the germination date, with the lowest RMSE of 12.97 days and a small bias of 2.04 days. In contrast, the threshold-based strategies of “5% of the amplitude” and “10% of the amplitude”, which are both generally used relative thresholds, exhibited poor performances. The “10% of the amplitude” strategy showed the largest error with an RMSE of 31.58 days and a bias of 22 days, indicating a significant overestimation of the germination date.

For the tillering stage, the time-windowed change detection method also showed a better performance than the threshold-based method. Similarly, “Point A” presented a low performance for retrieving tillering dates, but the time-windowed change detection method could reach better retrieval results. The temporal windows of 20 days, 25 days, and 30 days were applied. The “20 days after Point A” delivered the best performance, with the lowest RMSE of 4.41 days and a bias of −1.89 days. The “25 days after Point A” showed an RMSE of 5.06 days and a bias of 3.11 days, while the “30 days after Point A” method yielded an RMSE of 9.04 days and a bias of 8.11 days. In contrast, the threshold-based methods of “20% of the amplitude” and “30% of the amplitude” both exhibited higher errors.

For the elongation stage and maturity stage, the change detection method outperformed the threshold-based method. “Point B” showed high prediction accuracy in retrieving tillering dates, with an RMSE of 10.78 days and a bias of −3.16 days, while “50% of the amplitude (ascending)” presented lower accuracy with an RMSE of 23.47 days and a bias of 8.84 days. “Point E” showed a better performance in retrieving maturity dates, with an RMSE of 15.04 days and a bias of 7.41 days, while the “50% of the amplitude (descending)” method presented an RMSE of 18.50 days and a bias of 8.48 days.

Overall, the change detection method demonstrated higher accuracy than the threshold-based method in retrieving each sugarcane phenology stage. For the germination and tiller stages with sparse vegetation coverage, the change detection method, which was modified by a temporal buffer based on Point A, exhibited the highest retrieval accuracy. Point B and Point E corresponded to the stages of the fastest increases and decreases, respectively, showing high capabilities in retrieving the elongation and maturity stages. These results validated the effectiveness of the Double-Logistic function in capturing the key phases of the sugarcane growth process.

4.2. Assessment of Multiple Vegetation Indices for Phenology Retrieval

Based on the selected optimal retrieval strategies for each phenology stage, this study further evaluated the performances of various vegetation indices (the NDVI, EVI, EVI2, SAVI, RVI, GRVI) in retrieving the phenology across different growth stages (

Table 2. Performances of different vegetation indices for retrieving sugarcane phenology.

Vegetation Index	Germination	Tillering	Elongation	Maturity	Total
	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days
NDVI	12.97 (2.04)	4.41 (−1.89)	10.78 (−3.16)	15.04 (7.41)	12.48 (1.80)
EVI	17.71 (2.20)	11.65 (−6.89)	16.02 (−7.80)	17.41 (−4.88)	16.58 (−3.87)
EVI2	15.58 (1.75)	11.58 (−7.33)	13.56 (−6.58)	17.45 (−4.32)	15.24 (−3.54)
SAVI	15.40 (0.12)	12.34 (−8.89)	14.29 (−8.29)	16.13 (−1.72)	15.01 (−3.89)
RVI	26.99 (18.64)	24.21 (18.78)	22.03 (14.45)	19.34 (−8.70)	23.11 (9.17)
GRVI	35.12 (2.35)	43.66 (0.78)	20.05 (2.60)	26.71 (−2.81)	30.44 (0.67)

and Figure 7). A typical case is illustrated in Figure 8, which presents the time series of the various vegetation indices and the four phenological dates derived from them.

Figure 9 presents the accuracies of the phenology retrieved from the various vegetation indices. Among these vegetation indices, the NDVI obviously stood out in retrieving all four phenological dates, with an RMSE of 12.97 days and a bias of 2.04 days for germination, an RMSE of 4.41 days and a bias of −1.89 days for tillering, an RMSE of 10.78 days and a bias of −3.16 days for elongation, and an RMSE of 15.04 days and a bias of 7.41 days for maturity. As shown in Figure 8, the maturity date retrieved from the NDVI time series was relatively later than those derived from the other indices and showed a slight overestimation compared to those of the other indices, but it showed the highest estimation accuracy compared to the ground observation data. Overall, the NDVI achieved the highest accuracy, with a total RMSE of 12.48 days and a total bias of 1.80 days, while the SAVI followed, with a total RMSE of 15.01 days and a total bias of −3.89 days.

Compared to the NDVI and SAVI, the EVI and EVI2 demonstrated moderate performances, while the RVI and GRVI performed the worst in monitoring the sugarcane phenology. The EVI achieved a total accuracy with an RMSE of 16.58 days and a bias of −3.87 days. The EVI2, a simplified version of the EVI, showed higher accuracy than the EVI across all four phenological dates, achieving a total accuracy with an RMSE of 15.24 days and a bias of −3.54 days. As shown in Figure 8, the RVI responded more slowly to vegetation activity during the early growth stage and the vigorous stage compared to the other indices, resulting in the significant overestimation of the tillering and elongation stages. Conversely, the RVI responded more quickly during the maturity stage, achieving an RMSE of 19.34 days and a bias of −8.70 days, which is much better than the previous stages. The GRVI, which was sensitive to chlorophyll dynamics, exhibited the lowest retrieval accuracy throughout the entire growth cycle.

Based on these results, we can conclude that the NDVI, the widely adopted index, has the strongest ability to capture changes during the early growth, peak growth, and rapid aging stages, proving more effective at retrieving the germination, tillering, elongation, and maturity stages than the other adjusted vegetation indices.

4.3. Evaluation of the Spatiotemporal Fusion for Phenology Retrieval

To further evaluate the effect of spatiotemporal fusion on sugarcane phenology retrieval, the phenological dates retrieved from the unfused dataset were compared to those from the fused dataset. The unfused dataset was the coarse-resolution MODIS time series data, as presented in Section 2.2. The results are shown in

Table 3. Performance assessment of fused and unfused datasets for sugarcane phenology retrieval.

Dataset	Germination	Tillering	Elongation	Maturity	Total
	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days	RMSE (Bias)/Days
Fused	12.97 (2.04)	4.41 (−1.89)	10.78 (−3.16)	15.04 (7.41)	12.48 (1.80)
Unfused	18.25 (4.79)	8.19 (−0.5)	19.17 (−3.35)	19.02 (8.50)	17.97 (3.17)

. The fused dataset consistently outperformed the unfused dataset, demonstrating lower RMSE values across all stages. For the germination stage, the fused dataset achieved an RMSE of 12.97 days and a bias of 2.04 days, whereas the unfused dataset presented an RMSE of 18.25 days and a bias of 4.79 days. In the tillering stage, the fused dataset showed an RMSE of 4.41 days and a bias of −1.89 days, while the unfused dataset exhibited an RMSE of 8.19 days and a bias of −0.5 days, indicating significant overestimation and lower accuracy. Similarly, for the elongation stage retrieval, the fused dataset performed better with an RMSE of 10.78 days and a bias of −3.16 days, compared to the unfused dataset with an RMSE of 19.17 days and a bias of −3.35 days. In the maturity stage, the fused dataset achieved an RMSE of 15.04 days and a bias of 7.41 days, while the unfused dataset recorded a lower accuracy with an RMSE of 19.02 days and a bias of 8.50 days. The positive bias in the unfused dataset indicates a systematic error in the maturity stage retrieval.

Overall, the fused dataset demonstrated a superior performance in retrieving the phenological phases across all the growth stages compared to the unfused dataset. Spatiotemporal fusion using the ESTARFM significantly improved the monitoring accuracy in heterogeneous agricultural landscapes, reducing the RMSE by 21–46%. The unfused dataset exhibited larger errors, which severely limited its reliability for phenological phase retrieval.

5. Discussions

Prior studies on sugarcane phenology retrieval did not explore the relationship between remotely sensed signatures, and the corresponding ground-observed phenophases are unknown. This research bridges this critical gap by systematically evaluating different strategies and vegetation indices and investigating the effective retrieval methods for retrieving sugarcane’s key phenological dates.

The change detection method demonstrated a stronger performance in retrieving the key phenological dates, including germination, tillering, elongation, and maturity. Sugarcane exhibits gradual initial growth followed by rapid, explosive mid-season development. The third derivative effectively captures these instantaneous acceleration mutations. Previous studies directly linked inflection points with phenological transitions (e.g., Luo et al. retrieved crop emergence dates through inflection points [30]). However, sugarcane’s germination and tillering stages are characterized by low vegetation coverage, which prevents reliable detection via temporal curve features. The proposed time-windowed method realizes accurate retrieval and reduces uncertainties from soil background interference and temporal noise. For high-coverage stages (elongation and maturity), the key points along the curve retrieve the phenological dates successfully. In contrast, the threshold-based methods exhibit high amplitude dependence and noise sensitivity, which bring high uncertainties to the phenology retrieval.

Among the six vegetation indices evaluated, the NDVI demonstrated superior phenology retrieval accuracy throughout the entire sugarcane growth cycle. The secondary performers—the EVI, EVI2, and SAVI—exhibited substantial instability in large-scale, long-term applications. As illustrated in Figure 9, these adjusted indices showed extreme site-dependent variability: while achieving high accuracy at some locations, they produced significant overestimations or underestimations at others, compromising their overall reliability. The RVI displayed systematic overestimation during the germination, tillering, and elongation stages, proving that ratio-based formulations lack stability compared to the normalized difference indices. The GRVI performed the poorest, demonstrating that chlorophyll-sensitive visible bands alone are insufficient for phenology retrieval. This emphasizes the critical importance of near-infrared bands that probe internal leaf structures.

The limitation of MODIS time series data in sugarcane phenology retrieval stems fundamentally from spatial resolution mismatch. The ESTARFM fusion framework systematically addresses this through spatial decomposition (30 m resolution) and temporal interpolation (daily-scale reconstruction). This integrated approach significantly enhances the phenological retrieval accuracy, reducing the RMSE by 21–46% across key growth stages compared to MODIS-only methods in heterogeneous landscapes.

Although desirable effects were achieved for the proposed method regarding sugarcane phenology retrieval through remote sensing data, limitations still exist:

(1)

Retrieval algorithm scope

This study utilized change detection methods and threshold-based methods, which captured the mathematical metrics from vegetation index time series to retrieve phenological dates. As phenology retrieval is inherently dependent on high-fidelity time series reconstruction, inadequate reconstruction fidelity might propagate substantial errors in the derived phenological dates—particularly for rapid transition stages like germination onset. Current applications of machine learning algorithms in phenology retrieval remain underexplored. Future research should prioritize deep learning architectures (e.g., LSTM networks) to leverage fused multi-index temporal signatures. Furthermore, given the established linkage between crop phenology and climatic drivers—where the cumulative thermal time and hydrological conditions directly regulate growth rates—integrating meteorological variables (e.g., growing degree days, soil moisture) with remote sensing observations offers the potential to enhance the phenological retrieval accuracy.

(2)

Vegetation index selection

The current vegetation indices predominantly apply the visible to near-infrared spectral regions (400–900 nm) for basic photosynthetic activity monitoring. Expanding spectral analysis into the shortwave infrared (SWIR) region (1000–2500 nm)—through indices like the Normalized Difference Water Index (NDWI)—would further quantitatively track the crop moisture status [38]. Simultaneously, integrating radar-derived metrics would capture canopy structural dynamics, like stalk emergence and rapid elongation [39]. Furthermore, assimilating biophysical parameters (the FPAR, LAI) would bridge remote signatures with actual carbon allocation and growth processes [25].

(3)

Fusion model constraints

The ESTARFM operates under the assumption of constant land cover fractions throughout the observation period. This condition was satisfied in the current study due to the limited spatial extent of the monitored sugarcane sites and their stable land use patterns during the analysis timeframe. However, when scaling this approach to broader geographical regions—particularly those experiencing significant land use/cover changes (e.g., crop rotation, urban expansion, or forest disturbance)—the model may introduce significant uncertainties. In addition, the ESTARFM spatiotemporal fusion framework necessitates the acquisition of ≥2 cloud-free Landsat–MODIS image pairs per year from contrasting phenological seasons to accurately model the surface reflectance dynamics. This requirement poses substantial operational challenges in hyper-cloudy monsoonal regions like southeast China, where persistent cloud cover (>80% frequency during rainy seasons) often prevents the acquisition of valid images, inducing severe data gaps in the critical growth stages (e.g., tillering onset). To mitigate these limitations, alternative fusion frameworks or microwave-optical synergies should be explored to mitigate data gaps in high-precipitation zones. For instance, microwave-optical synergies leveraging the Sentinel-1 C-band SAR’s cloud-penetrating capability to anchor fusion during extended cloudy periods.

6. Conclusions

The present study focused on the retrieval of the sugarcane phenology and addressed three primary objectives. First, we identified the mathematical parameters corresponding to the field-observed phenological phases and developed optimized retrieval models for four key growth stages. Our analysis revealed that the change detection method effectively captured the sugarcane phenological dynamics, and the sparse canopy stage could be well retrieved using the developed time-windowed change detection method. Specifically, the best model for retrieving the germination date was identified as 25 days before Point A in the Double-Logistic curve, where Point A corresponded to the first local maximum point of the third derivative curve. For the tillering stage, the highest retrieval accuracy was achieved by 20 days after Point A. For the elongation and maturity stages, the best performance was associated with Point B and Point E, corresponding to the points with maximum slopes in the ascending and descending phases, respectively. Second, based on the acquired optimal retrieval strategies for each phase, we compared the performances of different vegetation indices. The results indicated that the NDVI was the most effective at retrieving the four key phenological stages—germination, tillering, elongation, and maturity—compared to the other adjusted vegetation indices. Third, it was demonstrated that in the heterogeneous agricultural landscapes of sugarcane plantations in South China, spatiotemporal fusion data significantly enhanced the retrieval performance. The fused dataset consistently outperformed the unfused dataset in the phenology retrieval across all growth stages. In conclusion, these findings contribute to more accurate and reliable methods for retrieving sugarcane phenology, providing valuable insights for improving agricultural monitoring practices.

In the future, several research efforts could be made to advance sugarcane phenological retrieval: the introduction of physiological indicators such as the FPAR and LAI to explore the sugarcane phenological dates, the development of machine learning-based retrieval algorithms driven by multiple indices instead of a single vegetation index, and the synergy of optical and radar data to eliminate the effects caused by frequent cloud cover. Additionally, further validation is necessary across a wider range of sugarcane observation sites in the future.