Mapping Rice Cropping Systems in Data-Scarce Regions Using NDVI Time-Series and Dynamic Time Warping Clustering: A Case Study of Maliana, Timor-Leste

Abstract

Mapping of rice-cropping regimes is crucial for effective irrigation planning and yield monitoring, particularly in data-scarce regions. We analyzed 48 months of 3 m PlanetScope NDVI data, aggregated to a 25 m hexagonal grid, and used Dynamic Time Warping Clustering to segment phenological patterns. Internal validation consistently identified two main clusters, indicating two dominant seasonality modes. Cluster 1 exhibited a higher mean NDVI, fewer low-canopy months, more vigorous growth periods, more peaks, and greater annual cycling, which suggests irrigated double cropping. Cluster 2 exhibited prolonged low NDVI values and a greater amplitude, consistent with single-rainfed systems. The rain–NDVI analysis supported these findings: Cluster 1 responded modestly to rainfall, whereas Cluster 2 exhibited a stronger and delayed response. Independent spatial checks confirmed these classifications. Off-season greenness, measured as NDVI above 0.50 from July to November, was concentrated near main and secondary canals and decreased with distance from intake points. This workflow combines DTW clustering with rainfall lag and off-season greenness analysis, effectively distinguishing between irrigated and rain-fed regimes using satellite time series. These findings are considered indicative rather than definitive, providing an assessment of cropping systems in Timor-Leste and demonstrating that DTW-based NDVI clustering offers a scalable approach in data-scarce regions.

1. Introduction

Rice (Oryza sativa L.) serves as the main staple food for over half of the global population, with more than 90% of its production and consumption occurring in Asia [1,2,3,4,5]. Timor-Leste remains a net importer of staple cereals, and rice production in particular still falls short of domestic demand, with national and international assessments indicating that imports routinely supply roughly half or more of the country’s rice requirements and a large share of total cereal consumption [6,7]. Local rice production consistently fails to meet demand, forcing the country to depend on imports and making it vulnerable to price and climate fluctuations, which complicate food security planning [6,8,9]. These structural challenges are exacerbated by inconsistent statistical coverage and fragmented monitoring systems, which limit the capacity for timely, evidence-based decision-making and underscore the need to strengthen agricultural data systems [10].

Within this national context, strategic irrigated rice basins that contribute disproportionately to domestic paddy output have been identified as priority targets for enhancing local productivity and reducing import dependence [11]. The Maliana basin in Bobonaro municipality serves as a prominent example. As a major irrigated rice plain, it combines gravity-fed irrigation schemes with peripheral fields that rely more heavily on rainfall and has been explicitly designated as a national “high-potential” rice zone [11,12]. The combination of high current and potential productivity, the coexistence of irrigated and rain-fed management, and partially documented infrastructure establishes Maliana as a policy-relevant and representative testbed for developing methods to map rice cropping systems in statistically under-documented environments using NDVI time series and Dynamic Time Warping (DTW) clustering [12].

Satellite Earth-observation time series provide a scalable approach to addressing information gaps by enabling repeated monitoring of vegetation dynamics across administrative boundaries [13,14]. The Normalized Difference Vegetation Index (NDVI), derived from red and near-infrared reflectance, serves as a physically interpretable indicator of vegetation canopy status and is widely utilized to monitor crop phenology under varying weather and cloud conditions [15,16,17]. For spatial aggregation, hexagonal grids are preferred over square grids because each cell has six equally spaced neighbors, thereby reducing edge effects and enhancing neighborhood representation in ecological and land-surface analyses [18]. In heterogeneous rice landscapes such as Maliana, where fields vary in water control, planting dates, and management intensity, integrating dense satellite time series, physically meaningful spectral indices, and spatial units that better capture neighborhood structure is especially valuable for characterizing spatial patterns in crop development [12].

When planting dates and crop cycles vary across regions, it is crucial to account for temporal misalignment before comparing time series. Dynamic Time Warping (DTW) offers a robust similarity measure by aligning sequences along the temporal axis. In the context of NDVI trajectories, DTW has been effectively employed to map rice cropping systems with flexible calendars, including national-scale applications in Vietnam [19]. Nevertheless, the application of DTW-based clustering at high spatial resolution, particularly when integrating neighborhood-preserving hexagonal units and explicit irrigation context, remains limited in small, data-scarce countries such as Timor-Leste, where traditional statistical sources are often sparse and outdated.

Unsupervised clustering frameworks require internal validation to determine an appropriate number of clusters and to assess their separation and compactness. Commonly used metrics include the Silhouette index, the Davies–Bouldin index, and the Calinski–Harabasz index [20,21,22]. In monsoon-affected agroecosystems, it is also important to link vegetation dynamics to hydroclimatic forcing. Because vegetation greenness typically responds to precipitation with lags ranging from weeks to months, rainfall–NDVI relationships are better analyzed as functions of time lag rather than purely synchronously. Many studies indicate that the highest correlations occur approximately one to two months after rainfall, and that the coupling between rainfall and NDVI differs between irrigated and rain-fed rice systems [23,24,25]. In rain-fed systems, a single peak in greenness often occurs about 40 to 50 days after rainfall [23,24,25]. These differences in lag structure are directly relevant to Maliana, where gravity-fed irrigation coexists with more rainfall-dependent plots along the basin margins [12].

This study presents a scalable workflow for characterizing rice phenology and cropping systems in data-scarce Timor-Leste by integrating 48 months of 3 m NDVI data with dynamic time warping (DTW)-based clustering on a hexagonal grid. The methodology (i) constructs continuous NDVI trajectories for each hexagonal cell without requiring cadastral inputs, (ii) assesses clustering quality using internal indices and a DTW-inertia elbow, and (iii) relates vegetation dynamics to hydroclimate through rainfall–NDVI cross-correlations across temporal lags. Threshold-based phenology metrics yield interpretable summaries of seasonal behavior, while infrastructure-aware and off-season greenness checks provide independent contextual validation [26]. The Maliana basin is selected as the focal area due to its concentration of rice production, the presence of functioning gravity-fed irrigation schemes alongside more rainfall-dependent zones, and its explicit designation as a national “high-potential” rice zone [11,12]. The objective is to produce indicative maps and diagnostics that clarify spatial patterns of rice management and probable water-control conditions, supporting discussions on irrigation scheduling, input allocation, and climate-risk planning within a transparent pipeline suitable for replication and scale-up in data-limited regions [12].

2. Study Area and Data Used

2.1. Brief Introduction to the Study Area

Maliana is located in the central region of Bobonaro municipality’s low-lying rice basin, an extensive area of alluvial clays at elevations below 100 m above sea level, fed by the Bulobu, Nunura, and Malibaka rivers. Its location, approximately 149 km southwest of Dili and a few kilometers from the Indonesian border, enables efficient movement of paddy to domestic urban markets and supports cross-border trade opportunities. The plain’s coordinates (8°52′–9°20′ S; 124°12′–125°26′ E) and its pronounced west–east topographic gradient form a natural catchment that accumulates water and fertile sediment, creating optimal conditions for irrigated rice cultivation [27]. In addition to these physical and hydrological benefits, Maliana is recognized as one of Timor-Leste’s principal irrigated rice basins and serves as a key production hub in national food security and rice-sector planning [6].

Production statistics confirm that Maliana serves as a key driver of Timor-Leste’s rice sector. From 2016 to 2019, the planted area increased from 866 ha to 3503 ha, and yields rose to 3.6 t ha⁻¹, significantly surpassing the national average of approximately 2.5 t ha⁻¹ [28]. Municipal records indicate that in 2024, Maliana produced 7173 t of unmilled rice, representing the largest contribution among Bobonaro’s post-administrative areas [6]. Both national and international evaluations consistently identify Maliana as one of only three “high-potential” paddy districts in Timor-Leste [29]. Agronomic surveys further reveal that about 1245 ha, or roughly 52% of Maliana’s cultivated land, are allocated to rice, while maize and other crops occupy the remainder, with typical yields of 3–4 t ha⁻¹ [7]. Community-level rice-security research also highlights Maliana, along with neighbouring Cailaco, as principal rice-producing sub-districts within Bobonaro, emphasizing the basin’s central importance to local food supply and livelihoods [11].

The current output relies on established hydraulic infrastructure. The twin gravity schemes, Maliana-I and Maliana-II (Figure 1), continue to supply perennial water to approximately 1800 hectares, despite flood damage that reduced the original command area from 2800 hectares. A JICA-funded upgrade of Maliana-I in 2008 increased its irrigable area from 600 to 1050 hectares. Collectively, these systems support a cropping intensity greater than 1.3 [30,31]. Concurrently, Maliana has been the focus of rice value-chain initiatives and best-practice demonstration projects led by the Government of Timor-Leste, JICA, and research partners, with the explicit objective of strengthening domestic rice production and reducing reliance on imported grain [12].

Maliana possesses a favorable hydro-geomorphic setting, demonstrated productivity, and effective hydraulic infrastructure within a landscape prioritized by national and international programmes as a key rice production zone [6,7,11,12]. These attributes establish the Maliana basin as a policy-relevant and representative testbed, justifying its selection as the primary study region for the development and evaluation of an NDVI time-series and Dynamic Time Warping (DTW) clustering workflow for rice cropping-system mapping in data-scarce environments (Figure 2).

2.2. Data Used

2.2.1. PlanetScope NDVI Time Series Imagery

PlanetScope multispectral scenes (PSScene—Surface Reflectance) from Planet Labs PBC (San Francisco, CA, USA) (http://www.planet.com/) were acquired for every clear-sky day between December 2018 and September 2022. Each CubeSat in the constellation captures four spectral bands—blue, green, red, and near-infrared (NIR)—at ≈3 m pixel size, providing the spatial detail needed to resolve individual paddy blocks and other smallholder fields.

All scenes are orthorectified and atmospherically corrected to surface reflectance, and are accompanied by quality metadata (acquisition time, sun/sensor geometry, cloud cover). These metadata were inspected to exclude images with more than 20% cloud coverage or severe off-nadir views.

For every retained scene, we computed the Normalized Difference Vegetation Index (NDVI) using

$${NDVI}_{PlanetScope}={\displaystyle \frac{{\rho NIR}_{PlanetScope}-{\rho Red}_{PlanetScope}}{{\rho NIR}_{PlanetScope}+{\rho Red}_{PlanetScope}}}$$

(1)

where $\rho NIR$ and $\rho Red$ represent surface reflectance values from the near-infrared and red spectral bands, respectively.

The Normalized Difference Vegetation Index (NDVI), derived from high-resolution PlanetScope imagery, offers significant advantages for monitoring rice crop development, health, and stress responses. The near-daily revisit frequency and meter-level spatial resolution of PlanetScope imagery enable detection of subtle and short-term changes in rice crop phenology. These phenological changes include critical stages such as transplanting, panicle initiation, and senescence [32]. NDVI is closely correlated with canopy chlorophyll content and structural variation, establishing it as a reliable indicator of rice crop vigor and physiological stress [33]. In rice production systems, NDVI time series capture dynamic phenological phases and environmental stresses, including flooding and drought [34]. Combining NDVI with other vegetation indices enables the discrimination of rice diseases at the sub-field scale, thereby enhancing its application in precision agriculture [35]. In this study, the NDVI rasters produced from PlanetScope imagery serve as the primary input time series for the phenological analysis and DTW-based clustering workflow.

2.2.2. Meteorological Data

Monthly average rainfall data were obtained from the Agro-Meteorology Section of the Ministry of Agriculture and Fisheries of Timor-Leste (https://timoragriresearch.weebly.com/weather-data.html, accessed on 2 August 2025). Rainfall measurements were recorded at 0.2 mm intervals, covering the same temporal extent as the PlanetScope imagery dataset (December 2018–November 2022). These meteorological data were used to analyze and validate the relationship between observed NDVI temporal patterns and local rainfall conditions, including the lagged response of vegetation greenness to precipitation.

2.2.3. Irrigation Canal and Orthophoto Data

Reference irrigation layers for the Maliana I scheme were obtained from the Japan International Cooperation Agency (JICA) Timor-Leste portal (https://www.jica.go.jp/english/overseas/easttimor/index.html, accessed on 2 August 2025). High-resolution orthophotos were sourced from the Centro Nacional de Dados Geoespaciais under the Ministry of Planning and Strategic Investment (https://mpie.gov.tl/pt/direcao-nacional-de-informacao-geoespacial/, accessed on 2 August 2025). All spatial datasets were harmonized to UTM WGS84 Zone 51S (EPSG:32751) and cross-checked against Google Satellite imagery.

Using QGIS 3.40.3 ‘Bratislava’, we digitized main, secondary, and tertiary canals and key control structures as GeoPackage layers, applied topology rules (no gaps/overlaps; class-consistent connectivity), and attributed each feature (class, status, source, width/lining where visible). The resulting canal vectors and structures were then overlaid on the DTW-derived NDVI clusters to interpret patterns of hydraulic reach and off-season greenness, providing spatial context for distinguishing between irrigated double-cropping and rain-fed regimes.

3. Methodology

Figure 3 Workflow for NDVI–DTW-based mapping of rice cropping systems in the Maliana basin. Multi-year PlanetScope imagery and monthly rainfall data are combined with orthophotos and irrigation canal information. After preprocessing, monthly NDVI composites are generated, limited to the mapped rice area, and aggregated over a 25 m hexagonal grid to create NDVI time series for each hexagon. These time series support the evaluation of candidate cluster numbers using DTW-based k-means, DTW inertia (Elbow), and internal validity indices (Silhouette, Calinski–Harabasz, Davies–Bouldin), leading to final DTW k-means clustering for the selected k. Phenology metrics and rainfall–NDVI lag profiles for each cluster enable agronomic labeling of clusters into cropping systems and probable water-supply regimes, which are then mapped spatially. Off-season NDVI mosaics and irrigation canal overlays are used to spatially validate the inferred cropping system classes.

3.1. Hexagon-Based Sampling and NDVI Time-Series Extraction

3.1.1. Hexagonal Grid Generation

A continuous lattice of hexagons was generated to construct spatially consistent sampling units across the entire study area in WGS 84/UTM Zone 51S (EPSG:32751). Each hexagon had an edge length of 25 m, which balanced spatial resolution with computational efficiency. The hexagonal configuration reduced edge effects and ensured six equidistant neighbors, thereby providing more uniform neighborhood representation than square grids and more effectively capturing curved gradients [36].

Each hexagon received a unique identifier (1–14,101). The grid was intersected with the cultivated area mask, and only hexagons that overlapped cultivated parcels were retained. Visual inspection confirmed that the 14,101 retained cells did not overlap with non-cultivated classes such as forest, urban or built-up areas, or water bodies. As a result, the analysis domain was restricted to cultivated parcels.

3.1.2. Scene-Wise NDVI Aggregation (Temporal Framework)

Cloud-screened PlanetScope NDVI rasters (Section 2.2.1) were aggregated at the hexagon level using batch zonal statistics. For each acquisition, the mean NDVI was calculated for every retained hexagon. Each processing run generated a table with the first column containing the hexagon ID and the remaining columns containing the mean NDVI values for each date.

All date-specific tables were merged so that each row represented a hexagon identifier and each column corresponded to an acquisition date, resulting in an N × T matrix (N = grid cells, T = monthly observations). This matrix provides a dense, quality-checked NDVI time series suitable for clustering, extraction of phenological metrics, and classification of cropping systems. Integrating hexagon-based spatial uniformity with batch zonal statistics for temporal consistency produces a high-resolution NDVI data cube that captures rice-field dynamics across spatial and temporal dimensions while minimizing contamination from non-agricultural pixels.

3.2. Time-Series Clustering Using DTW-Based K-Means

Dynamic Time Warping (DTW) facilitates non-linear temporal alignment of time series and serves as a foundation for phenological trajectory clustering in remote sensing applications [37,38]. The DTW-based k-means++ algorithm with barycenter averaging (DBA) enhances within-cluster coherence and consistently outperforms Euclidean-based methods [37,38]. DTW clustering has been applied to national-scale NDVI mapping of rice systems [19,39], crop and tree-cut detection using constrained DTW [40,41], and spatiotemporal mean-shift integration to differentiate temporally shifted or warped series [42,43]. Collectively, these methods illustrate the versatility of DTW clustering for analyzing complex vegetation indices.

DTW-based k-means clustering was applied to the monthly NDVI time series for each hexagon. The NDVI matrix was structured as an N × T array, where N represents the number of hexagons and T denotes the number of months. To ensure compatibility with tslearn’s DTW routines, which require an explicit feature dimension, the data were reshaped into an N × T × 1 array. In this format, the second axis corresponds to time, and the third axis serves as a per-timestep singleton feature [44]. This transformation retains all NDVI values and enables future extension to multivariate time series without altering the analytical pipeline.

Each time series was normalized to the [0, 1] range using Min–Max normalization [45], thereby prioritizing temporal patterns over absolute NDVI values. Subsequently, DTW-based k-means++ with DBA was applied to determine the candidate cluster number k within a specified range (Section 3.3). The DTW metric used a Sakoe–Chiba radius of 3, n_init of 3, max_iter_barycenter of 30, and random_state of 0. This configuration was selected to balance robustness and computational efficiency while ensuring reproducibility.

Geospatial processing and time-series clustering were performed on a Linux workstation (Ubuntu 20.04.6 LTS, Intel Core i9-10980XE CPU at 3.0 GHz with 36 cores, 252 GB RAM, NVIDIA Quadro RTX 6000 GPU, and 13 TB storage) using QGIS 3.40.3 and Python 3.13.5.

3.3. Determine the Optimal Value of k in K-Means Clustering

Clustering performance was evaluated and the optimal k was selected using a combination of DTW-inertia elbow analysis, dimensionality reduction with principal component analysis (PCA), and internal validity indices including Silhouette, Davies–Bouldin, and Calinski–Harabasz.

3.3.1. Elbow Method

For each candidate value of k between 2 and 8, the DTW-based k-means algorithm was applied to the normalized NDVI time series. The within-cluster sum of squared DTW distances, referred to as DTW inertia, was then calculated as follows:

$$WCSS={\sum }_{i=1}^k{\sum }_{j=1}^{n_i}distance {(x_j^{\left(i\right)},c_i)}^2$$

(2)

where ${({\mathit{x}}_{\mathit{j}}^{\left(\mathit{i}\right)},{\mathit{c}}_{\mathit{i}})}^2$ represents the distance between the j-th data point ${\mathit{x}}_{\mathit{j}}^{\left(\mathit{i}\right)}$ in cluster i and the centroid ${\mathit{c}}_{\mathit{i}}$ of that cluster [46].

The resulting WCSS values were plotted against k to identify an “elbow” point, beyond which additional clusters produced diminishing reductions in DTW inertia [46]. Given the subjectivity of visual elbow detection, this analysis was supplemented with quantitative criteria designed to detect significant changes in the WCSS decrease rate across k values [47,48]. This approach narrowed the plausible range of k values for further evaluation.

3.3.2. Silhouette Score and Principal Component Analysis (PCA)

Cluster separation and compactness in the reduced feature space were assessed using principal component analysis (PCA) of the normalized NDVI series. PCA consolidates variance into a limited set of orthogonal components, thereby reducing noise and collinearity in high-dimensional time-series data and enhancing the performance of k-means clustering [49,50].

The NDVI trajectories were projected onto the leading principal components, and Euclidean k-means clustering was conducted on the resulting PCA scores for various candidate k values. For each k, the Silhouette coefficient was calculated, defined for observation $i$ as;

$$s(i)={\displaystyle \frac{b\left(i\right)-a\left(i\right)}{\max \{a\left(i\right), b\left(i\right)\}}}$$

(3)

where $a\left(i\right)$ represents the mean dissimilarity between point $i$ and all other points within its own cluster, while $b\left(i\right)$ denotes the lowest mean dissimilarity between point $i$ and points in any other cluster [20]. Silhouette values approaching +1 indicate confident cluster assignment, values near 0 correspond to boundary cases, and negative values suggest possible misclassification [20].

Average Silhouette widths and Silhouette plots for k = 2 to 5 were analyzed alongside PCA scatterplots to assess cluster separation and to determine whether candidate k values identified by the DTW-inertia elbow corresponded with the structure of NDVI trajectories in the reduced space.

3.3.3. Cluster Validity Indices

To further validate the selection of k, the Davies–Bouldin Index (DBI) and the Calinski–Harabasz index (CH) were computed [51]. The DBI measures the average similarity between each cluster and its most similar cluster, defined as the ratio of within-cluster scatter to between-cluster centroid separation. Lower DBI values indicate more compact and well-separated clusters [52].

$$DBI={\displaystyle \frac{1}{K}}\sum _{i=1}^K{max}_{i\ne j} \left({\displaystyle \frac{s_i+s_j}{d_{ij}}}\right)$$

(4)

where K is the number of clusters, $s_i$ is the average distance of all samples in cluster $i$ from its centroid, and $d_{ij}$ is the Euclidean distance between the centroids of cluster $i$ and $j$ [52].

The Calinski–Harabasz index (CH) favors partitions with high between-cluster dispersion and low within-cluster dispersion; higher CH values reflect more distinct cluster structures [53,54].

$$CH(K)={\displaystyle \frac{B\left(K\right)·(N-K)}{W\left(K\right)·(K-1)}}$$

(5)

where $B\left(K\right)$ and $W\left(K\right)$ are the between- and within-cluster dispersions, $N$ is the number of samples, and $K$ is the number of clusters.

The final value of k was determined through consensus among the DTW-inertia elbow, Silhouette diagnostics, the Davies-Bouldin Index (DBI), and the Calinski-Harabasz (CH) criterion. Subsequently, DTW-based k-means clustering with the selected k was applied to the complete normalised NDVI dataset, and each hexagonal cell received a cluster label for subsequent mapping and interpretation.

3.4. NDVI and Rainfall Correlation

The timing and strength of the relationship between monthly NDVI and monthly rainfall for each cluster were quantified to relate vegetation dynamics to hydroclimatic forcing. Pearson cross-correlations $r_k=corr\left({rain}_t, {NDVI}_{t+k}\right)$ were calculated between the cluster-mean NDVI time series and the corresponding rainfall time series over lags k in the range [−6, +6] months. In this context, positive lags indicate that NDVI at time t is compared with rainfall at time t − k. Thus, positive lags represent greenness responses occurring k months after rainfall, consistent with established geophysical time-series conventions [55,56].

Time series were aligned over common months using pairwise complete observations to minimize spurious correlations arising from differing record lengths. Correlations were then evaluated across the entire lag window to capture typical soil-moisture and phenology response times at a monthly resolution [57,58]. For each cluster, the maximum positive-lag correlation (k ≥ 0) was reported as the interpretable delay between rainfall and vegetation response.

Lagged scatterplots of rainfall at time t − k versus NDVI at time t, incorporating ordinary least squares fits, were examined to verify the direction and approximate linearity of the relationship and to identify leverage or seasonality effects [57,59]. Given that both rainfall and NDVI exhibit seasonality and autocorrelation, cross-correlation peaks may indicate phase alignment in addition to causal responses. Consequently, interpretation focused on the positive-lag side and was corroborated using these scatterplot diagnostics [56,60].

4. Results

4.1. Optimal Number of Clusters

The elbow analysis of DTW-inertia indicates that the curve lies between k = 3 and k = 4. Marginal gains in DTW inertia are 7.8% from k = 2 to k = 3 and 4.2% from k = 3 to k = 4. For k values greater than 4, improvements decrease to approximately 3.0%, 2.5%, 1.5%, and 1.0% for k = 5 to k = 8. Based on the marginal-gain criterion, the elbow occurs at k = 4. Beyond this point, the curve flattens approximately linearly, indicating diminishing returns at higher values of k, as illustrated in Figure 4. Internal indices support this heuristic, as shown in

Table 1. Silhouettes score for k = 2 to k = 5.

k	Silhouette
2	0.517
3	0.472
4	0.477
5	0.447

and Figure 5. The mean silhouette value reaches a maximum of 0.517 at k = 2, indicating two distinct clusters with minimal overlap. The silhouette distribution for Cluster 1 displays a broad, high-density band, while Cluster 0 exhibits only a minor negative tail. When k increases from 3 to 5, these groups are further subdivided, but the mean silhouette values decrease to 0.472, 0.477, and 0.447, respectively. This decline reflects reduced cluster separation and cohesion [20]. Greater principal component analysis (PCA) overlap, narrower bars, and longer negative tails in the silhouette plots further demonstrate diminished internal consistency [20].

Additional validity metrics support this conclusion. The Davies–Bouldin Index reaches its minimum at 1.632 for k = 2 and increases for higher k values (1.721, 1.779, 2.010 for k = 3 to 5). In contrast, the Calinski–Harabasz Index is maximized at 4231.683 for k = 2 and decreases as k increases (3544.408, 3135.728, 2561.544), as shown in

Table 2. Cluster Validity Indices for k = 2 to k = 5.

k	Davies-Bouldin	Calinski-Harabasz
2	1.632	4231.683
3	1.721	3544.408
4	1.779	3135.728
5	2.010	2561.544

. Collectively, these quantitative and qualitative diagnostics indicate that k = 2 is the statistically and substantively optimal solution, preserving the dominant phenological structure required to distinguish the principal rice-cropping regimes, despite the elbow’s steepest decline spanning k = 2 to 3. Based on this selection, cluster-wise NDVI and rainfall time series are summarized for interpretability in Figure 6.

4.2. Crop Phenology Metrics

The 48-month PlanetScope NDVI record was analyzed using operational vigor thresholds: low canopy (NDVI < 0.45), vigorous peak (local maximum with NDVI ≥ 0.55), and vigorous month (NDVI ≥ 0.60). These thresholds identify periods of sparse canopy, peak greenness, and sustained vigor [61,62]. Cluster 1 shows higher mean greenness, fewer months with low canopy cover, and more months exceeding vigor thresholds than Cluster 2. This indicates that Cluster 1 has longer and more frequently renewed growing periods. Although both clusters reach similar maximum NDVI values, Cluster 2 maintains low NDVI for longer and exhibits greater seasonal amplitude, reflecting more pronounced off-season intervals.

Table 3. Crop phenology metrics.

Metric	Period	Cluster 1	Cluster 2
Average NDVI		0.5420	0.4693
Max		0.7375	0.7244
Min		0.3555	0.2962
Amplitude (max–min)		0.3819	0.4283
Cycles/year		2	1.5
No. of months < 0.45	(Dec18–Nov22)	6	25
	(Dec18–Nov19)	2	6
	(Dec19–Nov20)	1	6
	(Dec20–Nov21)	2	6
	(Dec21–Nov22)	1	7
No. of months > 0.60	(Dec18–Nov22)	11	7
	(Dec18–Nov19)	1	1
	(Dec19–Nov20)	1	2
	(Dec20–Nov21)	6	2
	(Dec21–Nov22)	3	2
Peaks > 0.55 NDVI	(Dec18–Nov22)	8	6
	(Dec18–Nov19)	2	1
	(Dec19–Nov20)	2	2
	(Dec20–Nov21)	2	2
	(Dec21–Nov22)	2	1

provides detailed values, including the average, minimum, and maximum NDVI; amplitude; counts of months below 0.45 and above 0.60; peak counts above 0.55; and cycles per year. These results are further validated by the off-season NDVI spatial distribution map, which visually relates cluster results to NDVI spatial patterns.

4.3. NDVI-Rainfall Correlation

Positive-lag peaks indicate that rainfall events precede increases in NDVI. Cluster 1 exhibits a rapid response, with a maximum correlation coefficient of approximately 0.36 at a one-month lag. In contrast, Cluster 2 displays a stronger but delayed response, reaching a maximum correlation of approximately 0.60 at a three-month lag. These results correspond to Figure 7 and

Table 4. Monthly rainfall–NDVI Pearson cross-correlations by lag k for each cluster.

Lag (k)	Cluster 1 (Pearson r)	Cluster 2 (Pearson r)
−6	−0.332	−0.268
−5	−0.269	−0.348
−4	−0.116	−0.593
−3	0.076	−0.600
−2	0.109	−0.548
−1	0.324	−0.325
0	0.296	0.005
+1	0.358	0.314
+2	0.259	0.525
+3	0.063	0.595
+4	−0.118	0.483
+5	−0.262	0.230
+6	−0.332	−0.097

and are consistent with the typical one- to three-month vegetation response to moisture recharge observed in monsoon-affected systems [63].

At negative lags, correlations are predominantly negative, indicating low greenness prior to rainfall and a decline in vegetation following the conclusion of the rainy season [64]. The shorter and weaker lag observed in Cluster 1 is consistent with buffered conditions, such as irrigation or residual soil moisture. Conversely, the longer and stronger lag in Cluster 2 suggests rain-dependent growth that requires moisture accumulation before canopy expansion [65]. Table 4 presents the complete lag profiles, and Figure 7 displays the correlation curves.

4.4. Cropping System Assignment

DTW clusters are translated into agronomic classes by integrating the phenology summary in Table 3 with rainfall–NDVI lag behavior (Figure 7; Table 4). DTW differentiates time series based on the shape of seasonal trajectories and is established for rice-calendar mapping in Southeast Asia [19]. Standard diagnostics such as seasonal amplitude, within-year peaks, duration above vigor thresholds, and cycles per year reveal clear contrasts (see Table 3). Thresholds around NDVI values of 0.5 to 0.6 indicate sustained, dense canopy during peak rice growth [61,62].

The annual cycle frequency was determined by dividing the number of NDVI peaks exceeding 0.55 by the observation period (in years) for each cell, then averaging across clusters. For example, a value of 1.5 cycles per year in Cluster 2 suggests either a primarily single rain-fed season with an intermittent second season in some parcels or years, or a mixture of single- and occasional double-season parcels within the cluster.

In monsoon systems, vegetation typically responds within one to three months after rainfall events [66]. Cluster 1 shows a short, damped positive-lag peak, suggesting buffered water availability from irrigation or residual soil moisture. Therefore, Cluster 1 is identified as an irrigated double-cropping system. In contrast, Cluster 2 displays a stronger, delayed peak, indicating rain-dependent growth that requires moisture accumulation; accordingly, Cluster 2 is identified as a single rain-fed cropping system [36,61,62,64,65,66]. Figure 8 illustrates the spatial distribution of these cropping systems.

4.5. Visual Validation with Off-Season NDVI and Irrigation Infrastructure

PlanetScope (3 m) off-season NDVI mosaics (June to November) in Figure 9 show two distinct regimes corresponding to the DTW clustering. Double-cropping fields maintain moderate to high canopy cover throughout most of the dry season, with large areas remaining above typical vigor thresholds. In contrast, single-cropping fields dry down soon after harvest, indicating fallow or sparse cover from July onward. This visual distinction supports the phenology differences summarized in Table 3, with Cluster 1 showing more sustained greenness and frequent high-NDVI peaks compared to Cluster 2. These findings confirm a consistently intensified double-cropping calendar versus a predominantly single-season system.

Overlaying the irrigation canal networks clarifies the underlying mechanism. Persistent off-season greenness is associated with the upgraded main canal and denser lateral canals, but this effect diminishes with greater distance from the intake and toward canal terminal, as shown in Figure 10. Reported dry-season design discharges, including approximately 1.37 to 1.35 m³/s at the headworks, 0.96 to 0.16 m³/s along stronger secondary canals near Ramascora, and 0.39 to 0.17 m³/s near Ritabou, correspond to this spatial pattern [67]. Where conveyance capacity and delivery reliability are sufficient, fields remain green and support double-cropping. In contrast, where capacity decreases or losses accumulate with distance, fields dry and support only single cropping.

The integration of off-season mosaics and canal overlays offers map-based validation. Cluster 1 aligns with irrigated double-cropping within the hydraulic reach of the canal network, whereas Cluster 2 represents rain-fed areas where conveyance is limited or distant. These findings indicate that the phenological signal is directly associated with the physical infrastructure and its effective service area.

Taken together, the DTW clusters, phenology metrics, rainfall-lag profiles, and off-season NDVI patterns act as a remote diagnostic of water-supply conditions. Fields assigned to the double-cropping cluster and showing persistent off-season greenness within the canal command area can be interpreted as having reliable irrigation, whereas fields in the single-cropping cluster with strong rainfall dependence represent rain-fed or marginally irrigated plots. In data-scarce settings where canal condition and service areas are poorly documented, this classification provides a first-order, remotely sensed proxy for water-supply type at field scale.

5. Discussion

PlanetScope NDVI time series, using DTW-based k-means clustering on a 25 m hexagonal grid, identified two main phenological regimes in the Maliana basin: an irrigated, often double-cropping system and a predominantly single-season, rainfed system. These regimes differ in mean NDVI, duration of vigorous canopy cover, peak frequency, and seasonal amplitude, as shown by phenology metrics and cycle counts in Table 3. The irrigated regime sustains higher greenness for longer periods, with more months above vigor thresholds and more cycles per year. In contrast, the rainfed regime shows extended low NDVI and more pronounced off-season intervals. These differences align with rainfall–NDVI lag profiles: the irrigated cluster has a shorter, weaker positive lag, while the rainfed cluster displays a stronger, delayed response, indicative of buffered versus rainfall-dependent water-supply conditions.

The integration of rainfall timing with irrigation infrastructure provides additional context for interpreting these regimes. Off-season NDVI mosaics reveal persistent canopy cover near main and secondary canals, which decreases with greater distance from intake structures and reduced conveyance capacity. This spatial pattern aligns with reported dry-season discharges and the mapped canal hierarchy, supporting the conclusion that the high-vigor cluster comprises fields with more reliable access to gravity-fed irrigation. Conversely, areas exhibiting the low-vigor, single-cycle regime are generally situated at the periphery of the hydraulic network or beyond its effective reach, where rainfall and soil moisture accumulation primarily determine growth. Together, the DTW clusters, phenology metrics, rainfall–NDVI lags, and off-season greenness patterns form a coherent remote diagnostic of relative water control and cropping intensity, although they do not constitute a fully validated classification of management practices.

The results directly address the study objectives. First, constructing a multi-year, high-resolution NDVI time series on a boundary-independent hexagonal grid produces stable trajectories for each cell and reduces edge artefacts, thereby meeting the requirement for spatial units appropriate for phenological analysis. Second, the derived metrics, including threshold-based vigor durations, peak counts, amplitudes, and cycles per year, provide concise and interpretable descriptors of rice phenology that capture both seasonal timing and canopy vigor. Third, rainfall–NDVI cross-correlations and lag structures establish links between these phenological indicators and hydroclimatic forcing, enabling differentiation between buffered and rain-dependent regimes. Finally, the resulting cluster maps, together with summary statistics and off-season NDVI checks, constitute decision-oriented products that can inform irrigation scheduling, canal maintenance prioritization, and identification of rainfall-sensitive zones. However, given the current level of validation, these products should be regarded as screening tools rather than operational decision rules.

Future research should emphasize field-based validation and integration of multiple data sources. Conducting ground surveys to collect plot-level calendars, irrigation schedules, and yield records would enable a more rigorous evaluation of the correspondence between DTW-derived clusters and actual cropping systems and performance outcomes. Incorporating additional spatial data, such as detailed canal-condition surveys, soil maps, topographic indices, and high-frequency meteorological observations, would facilitate attribution of observed NDVI differences to specific mechanisms and allow for quantification of uncertainty in regime assignment. From a remote sensing perspective, integrating PlanetScope NDVI with Synthetic Aperture Radar (SAR) and complementary spectral indices could address cloud and saturation limitations and enhance sensitivity to critical stages, including flooding and transplanting. These advancements would transition the workflow from a data-driven diagnostic of phenological regimes to a calibrated system capable of monitoring and cautiously predicting rice-sector behavior under climate variability and infrastructure change.

Transferability and Use in Data-Limited Regions

The workflow developed in this study is descriptive rather than predictive. It delineates phenological regimes, relates them to rainfall and infrastructure, and provides maps of probable water-supply conditions; however, it does not forecast yields or end-of-season outcomes. The stability of the two-cluster solution over four years and the consistent differences in rainfall–NDVI lag structures indicate that the Maliana centroids can serve as canonical trajectories for irrigated and rainfed rice under current conditions. NDVI time series from subsequent years or from neighboring basins could, in principle, be elastically aligned to these centroids early in the season. This approach would allow new cells to be provisionally assigned to an irrigated-like or rainfed-like regime as their partial trajectories develop.

In data-limited regions, early-season centroid-based assignment may support exploratory mapping of likely double- versus single-cropping zones or highlight areas where evolving trajectories diverge from established patterns. However, without plot-level data on yields, management, and water delivery, the predictive skill of these applications remains unquantified. Therefore, any use of the current centroids for forecasting should be considered probabilistic and exploratory. Such applications are appropriate for generating hypotheses, identifying anomalies, or guiding targeted field inspections, rather than for direct operational decision-making. As additional years of imagery and ground observations become available, the centroids and their associated uncertainty can be updated, enabling the development of more formal, data-driven priors for regional crop-monitoring and risk-assessment systems.

6. Conclusions

The integration of high-resolution PlanetScope NDVI, dynamic time warping (DTW)-based k-means clustering, and a 25 m hexagonal grid enables the extraction of two operationally meaningful rice phenological regimes in a data-scarce basin. Internal validity indices, including Elbow, Silhouette, Davies–Bouldin, and Calinski–Harabasz, consistently support a two-cluster solution. Cluster-specific phenology metrics and rainfall–NDVI lag profiles differentiate an irrigated, frequently double-cropping regime from a predominantly single-season, rainfed regime in Maliana. These classes are characterized by distinct NDVI amplitudes, durations of vigorous canopy, cropping-cycle frequencies, and moisture-response lags. Furthermore, they spatially correspond with the documented hydraulic reach of primary and secondary irrigation canals as well as with off-season greenness patterns.

In addition to generating the first high-resolution, multi-year map of rice cropping regimes for a priority basin in Timor-Leste, this workflow provides a replicable template for application in other data-limited regions. The approach utilizes widely available satellite imagery, basic rainfall time series, and a boundary-independent hexagonal grid, making it appropriate for basins and countries with incomplete or outdated agronomic statistics and irrigation inventories. The Maliana centroids and associated metrics offer quantitative reference trajectories for irrigated and rainfed rice regimes, supporting data gap-filling and enabling comparative analyses of cropping intensity and water control across basins, provided that careful local calibration is conducted.

The primary limitation of this study is the lack of plot-level agronomic and yield data, as well as detailed records of on-farm water delivery and management practices. This limitation prevents rigorous external validation and necessitates interpreting the mapped regimes as indicative rather than definitive representations of cropping systems and performance. However, the observed internal consistency among phenological metrics, rainfall–NDVI lags, and irrigation infrastructure suggests that the workflow provides valuable first-order diagnostics of relative water supply and cropping intensity in smallholder rice systems. Future research should integrate this framework with coordinated field campaigns, crop calendars, yield surveys, and multi-sensor remote sensing to calibrate predictive skill and develop robust tools for rice-sector monitoring and climate-risk planning in Timor-Leste and other data-scarce regions.