Mapping Multi-Crop Cropland Abandonment in Conflict-Affected Ukraine Based on MODIS Time Series Analysis

Abstract

Since the outbreak of the Russia–Ukraine conflict in 2022, Ukraine’s agricultural production has faced significant disruption, leading to widespread cropland abandonment. These croplands were abandoned at different stages, primarily due to war-related destruction and displacement of people. Existing methods for detecting abandoned cropland fail to account for crop type differences and distinguish abandonment stages, leading to inaccuracies. Therefore, this study proposes a novel framework combining crop-type classification with the Bias-weighted Time-Weighted Dynamic Time Warping (BTWDTW) method, distinguishing between sowing and harvest abandonment. Additionally, the proposed framework improves accuracy by integrating a more nuanced analysis of crop-specific patterns, thus offering more precise insights into abandonment dynamics. The overall accuracy of the proposed method reached 88.9%. The results reveal a V-shaped trajectory of cropland abandonment, with abandoned areas increasing from 28,184 km² in 2022 to 33,278 km² in 2024, with 2023 showing an abandoned area of 24,007.65 km². Spatially, about 70% of sowing abandonment occurred in high-conflict areas, with hotspots of unplanted abandonment shifting from southern Ukraine to the northeast, while unharvested abandonment was observed across the entire country. Significant variations were found across crop types, with maize experiencing the highest rate of unharvested abandonment, while wheat exhibited a more balanced pattern of sowing and harvest losses. The proposed method and results provide valuable insights for post-conflict agricultural recovery and decision-making in recovery planning.

1. Introduction

Cropland, as the foundation of the global food supply, is increasingly threatened by population growth and socio-economic pressures [1,2,3]. To meet the needs of an estimated 9.7 billion people by 2050, food production must rise by 70% [4]. However, cropland loss is accelerating due to climate change, policy shifts, and especially armed conflict—the most disruptive factor—causing soil degradation, landmine contamination, and severe threats to ecosystems and food security [5,6].

Ukraine, long regarded as a major global food exporter, possesses some of the world’s most fertile soils. However, since February 2022, large-scale armed conflict between Russia and Ukraine has severely disrupted agricultural production. Official data from Ukraine’s Ministry of Agrarian Policy show a 27% drop in sown area, a 30.4% decline in grain output, and a 15.5% decrease in agricultural exports in 2022.

Among the many impacts of the war, widespread cropland abandonment is one of the most visible and urgent issues [7], particularly in frontline oblasts such as Donetsk, Luhansk, Zaporizhzhia, and Kherson [8]. Estimates suggest that over 8 million hectares of farmland in Ukraine may be affected by landmines and war debris, posing long-term risks even after hostilities cease [9]. In this context, mapping and quantifying cropland abandonment across different crop types is essential for assessing war impacts and informing recovery strategies.

In recent years, an increasing number of studies have focused on monitoring cropland distribution and agricultural activities during and after wartime. For example, 30 m resolution remote sensing imagery was used to study agricultural changes following the Syrian civil war, showing increased interannual variation and spatial heterogeneity in croplands [10]. Similarly, between 2016 and 2018, remote sensing data revealed a 16% reduction in cropland area in South Sudan, with conflict-affected areas experiencing a larger decrease of 22.9% compared to 6.5% in non-conflict areas [11]. Dual-period NDVI comparisons were applied to delineate abandonment in Ukraine before and after the conflict, highlighting its expansion in the east and south [12]. The TWDTW algorithm was applied to Sentinel data to map abandoned land in Ukraine’s eastern oblasts, with cumulative NDVI proposed as a proxy for yield loss [8]. A recent study used machine learning and Sentinel data to assess agricultural land use changes in Ukraine from 2016 to 2024, revealing a 10% reduction in arable land, particularly in conflict zones [13]. Additionally, other studies have integrated GIS and machine learning techniques to assess Ukraine’s agricultural damage and food insecurity [14,15]. While these studies provide valuable insights, most are spatially limited (focusing primarily on eastern Ukraine), temporally constrained (typically covering the period from 2022 to 2023), and treat cropland as a homogeneous category. More critically, few distinguish between two fundamentally different types of abandonment—sowing abandonment (fields left unplanted) and harvest abandonment (fields planted but not harvested)—which reflect distinct stages of agricultural disruption and are driven by different underlying factors.

For the purpose of this study, abandoned cropland is defined as land that has not been cultivated for at least one year due to the ongoing conflict in Ukraine. This definition differs from the FAO’s traditional criteria, which defines abandonment as land that has not been planted for more than five years. Our definition reflects the unique circumstances of conflict-induced disruption, where lands may remain uncultivated for shorter periods due to the immediate effects of war. In normal years, distinguishing between rotational fallow and true abandonment can be challenging. However, in the context of armed conflict, rotational fallow is not a valid explanation, as it relies on organized agricultural management, which is disrupted by war. Farmers are unable to manage land for rotational fallow, and agricultural activities are halted, leading to true abandonment rather than planned rest periods.

Existing approaches to extracting abandoned cropland information include three primary methods [12]: direct land feature classification [16,17], multi-period comparative analysis [18,19], and change detection based on time series remote sensing indices [20,21,22]. Among these, time series NDVI methods are particularly effective in identifying the onset, duration, and extent of abandonment, outperforming snapshot-based techniques in terms of accuracy. Although high-resolution data from Landsat and Sentinel satellites offer detailed spatial information, their utility is often limited by cloud cover and infrequent revisits [19]. In contrast, MODIS NDVI products, with their higher temporal resolution and consistent data quality, are better suited for large-scale, long-term monitoring of vegetation changes and cropland abandonment like that in Ukraine [23,24].

This study aims to address current gaps by conducting a nationwide analysis of cropland abandonment in Ukraine from 2022 to 2024, using MODIS time series NDVI data. A refined remote sensing framework is proposed that integrates crop-type classification with a phenology-based time series similarity approach. Within this framework, cropland abandonment is defined as land that ceased active use due to conflict-related disruption and is further categorized into two subtypes. This classification allows for more precise identification of abandonment drivers and enables analysis of crop-specific patterns under wartime stress. Specifically, this study contributes to the existing body of knowledge by enhancing abandonment detection through the development of a Bias-weighted Time-Weighted Dynamic Time Warping (BTWDTW) method. Compared to traditional time series approaches, BTWDTW incorporates both the phenological alignment and asymmetric weighting of residuals, enabling finer differentiation between sowing and harvest abandonment. This advancement improves temporal sensitivity and classification accuracy, particularly in complex, conflict-affected environments. The study also quantifies abandonment trends by crop type, and explores their association with conflict zones. The findings provide scientific evidence for understanding the agricultural consequences of war and offer insights for post-conflict recovery planning and food system resilience in crisis-affected regions.

2. Materials

2.1. Study Area

Ukraine, located in Eastern Europe, covers approximately 603,700 square kilometers and is administratively divided into 24 oblasts and one autonomous republic, Crimea [8,12,25], as shown in Figure 1. It has abundant agricultural resources, with about 427,000 square kilometers of agricultural land—nearly 70% of its territory—including 325,000 square kilometers of arable land, accounting for 27% of Europe’s total [26]. Before the Russia–Ukraine conflict, Ukraine was a leading global exporter of sunflower oil, wheat, and maize, with agriculture contributing over 15% of its GDP [27,28]. As the “breadbasket of Europe,” it played a vital role in international grain and oilseed trade [29,30].

The Russia–Ukraine conflict has predominantly affected southeastern regions, including Luhansk, Donetsk, Zaporizhzhia, Kherson, Crimea, Kharkiv, and Dnipropetrovsk. As shown in Figure 1, based on international military assessments and LIVEUAMAP data (https://liveuamap.com (accessed on 27 March 2025)), approximately 38% of Ukraine’s total cropland lies within these conflict-affected zones [14,15].

2.2. Data Sources

This study employed the MODIS vegetation index product (MOD13Q1), which provides NDVI and EVI data at a spatial resolution of 250 m, synthesized every 16 days using the Maximum Value Composite (MVC) method, along with quality control information. In addition, the MODIS surface reflectance product (MYD09A1), with a 500 m spatial resolution and an 8-day temporal interval, was used to calculate two additional spectral indices related to vegetation moisture and pigment dynamics. The MODIS global annual land cover product (MCD12Q1) was further used to extract the spatial distribution of cropland in Ukraine. Sentinel-2 imagery offers high spatial resolution; however, over large areas, cloud cover and shadow effects frequently lead to data gaps, which limit the continuous acquisition of complete time series observations. As a result, Sentinel-2 data were primarily used to optimize training samples, assess classification accuracy, and support a comparative analysis of small-scale abandoned cropland. Administrative boundary data for Ukraine were obtained from the GAUL (Global Administrative Unit Layers) dataset maintained by the FAO. All remote sensing data were downloaded from the Google Earth Engine (GEE) platform [31].

2.3. Data Preprocessing

The MODIS Vegetation Indices (MOD13Q1) from 2019 to 2024 were temporally interpolated to an 8-day interval. Surface reflectance data from MODIS (MYD09A1) for the years 2019–2021 were used to calculate the Land Surface Water Index (LSWI) [32] and the Normalized Difference Yellow Index (NDYI) [33], which were then resampled to a spatial resolution of 250 m. This ensured consistency across all four indices (NDVI, EVI, LSWI, NDYI) in both spatial and temporal dimensions (250 m, 8-day). LSWI and NDYI were derived using the following formulas:

$$LSWI={\displaystyle \frac{\rho NIR-\rho SWIR1}{\rho NIR+\rho SWIR1}}$$

(1)

$$NDYI={\displaystyle \frac{\rho Green-\rho Blue}{\rho Green+\rho Blue}}$$

(2)

where ρ represents surface reflectance in the corresponding spectral bands from MODIS. The LSWI is sensitive to vegetation water content, whereas NDYI captures crop yellowing signals indicative of blossom or senescence.

Subsequently, the IGBP land cover classification band from the MODIS Land Cover Product was used to mask non-cropland areas. The dataset was then clipped to the national boundaries of Ukraine using administrative vector data from the FAO GAUL dataset.

2.4. Validation Sample Preparation

In this study, a unified set of multi-label sample points was constructed to validate both crop classification and cropland abandonment detection. Given the resolution difference between MODIS data (250 m for MOD13Q1 and 500 m for MYD09A1) and Sentinel-2 data, we carefully selected large cropland patches in the Sentinel-2 imagery that covered areas equivalent to at least four 250 m MODIS pixels to ensure alignment with the MODIS data resolution. In selecting sample points, we also ensured that the points were spaced sufficiently apart to maintain representativeness. This procedure minimized the potential issues caused by resolution disparity and allowed for a more accurate analysis of abandoned cropland. Each point includes the following: (1) the dominant crop type during the pre-conflict period (2019–2021), and (2) a label indicating abandonment status during 2022–2024, based on changes in crop growth. These spatially consistent, temporally annotated samples support both classification and abandonment analysis.

A total of 2845 classification points covering six crop types—wheat, sunflower, rapeseed, maize, soybean, and potato—were compiled from Chen et al. [15], and further validated using high-resolution imagery (Esri World Image, Sentinel-2), EUCROPMAP [34], and the crop map produced by the Kyiv Polytechnic Institute (KPI) team (https://ukraine-cropmaps.com (accessed on 5 November 2024)) [35,36] via visual interpretation.

For abandonment labeling, each point was classified as either “sowed” or “unsowed” using a visual inspection tool on GEE in conjunction with predefined seasonal NDVI thresholds (Figure 2). These thresholds are typically determined based on the crop’s lifecycle and growth stages. When NDVI increases from low values (close to 0) and exceeds 0.4 during the maturation period, and remote sensing imagery shows clear vegetation coverage, the area is labeled as “sowed.” Conversely, if NDVI shows little to no increase throughout the growing season, remaining below 0.4 during maturation, and there is no visible vegetation coverage in the imagery—indicating bare soil—the area is classified as “unsowed”.

3. Methods

3.1. Overview of Research Framework

This study developed a comprehensive methodology to map and analyze cropland abandonment in Ukraine amid ongoing conflict (Figure 3). First, crop type classification was performed using three years of pre-war MODIS time series data (NDVI, EVI, LSWI, NDYI). A dynamic segmentation module was then applied to define crop growing and harvesting periods. Abandonment was identified by distinguishing between sowing and harvest abandonment, using the BTWDTW algorithm to compare field NDVI profiles with reference curves. Accuracy was evaluated using ground truth data and benchmarked against traditional methods. Finally, spatiotemporal patterns of abandonment were analyzed to reveal the conflict’s evolving impacts on agriculture across different years and regions.

3.2. Crop Type Classification

Crop type classification was conducted using the RandOm Convolutional KErnel Transform (ROCKET) [37] algorithm, a fast and widely adopted method for multivariate time series classification [38]. The algorithm transforms raw time series data into a high-dimensional feature space through the application of a large number of randomly generated 1D convolutional kernels.

The overall pipeline of the ROCKET algorithm is summarized in Figure 4. Multiple such kernels are applied in parallel to the input time series, and the resulting MAX pooling and PPV (Proportion of Positive Values) features from each kernel are concatenated to form a comprehensive feature vector. This vector is then used to train a linear classifier, specifically Ridge Regression, for the final prediction.

The input features consisted of a multivariate time series derived from four vegetation indices—NDVI, EVI, LSWI, and NDYI—extracted from MODIS products and resampled to a temporal resolution of 8 days and a spatial resolution of 250 m. Classification was conducted on a per-pixel basis for each year, resulting in annual crop type maps for 2019, 2020, and 2021.

3.3. Period Segmentation

Accurate segmentation of crop growth and harvest periods is key to identifying abandoned planting or harvesting. Typically, abandoned fields show low, flat NDVI curves, while cultivated fields exhibit seasonal variation with a peak followed by a decline at harvest. Harvest abandonment may mimic normal growth but lacks the expected post-maturity drop.

To improve detection accuracy, this study introduces a dynamic segmentation module based on NDVI curves and Ukraine’s crop calendar. Rather than using fixed dates, the module identifies field-specific maturity peaks within a general window (Julian Day 97–257) and defines growth and harvest periods relative to this point. Crop types are grouped into autumn (e.g., wheat, rapeseed) and spring (e.g., maize, sunflower), with respective period lengths of 56 and 64 days. The harvest window matches the growth duration, accommodating phenological variation across regions and crops (Figure 5).

3.4. Bias-Weighted Time-Weighted Dynamic Time Warping (BTWDTW)

Traditional Dynamic Time Warping (DTW) and its time-weighted variant Time-Weighted Dynamic Time Warping (TWDTW) have proven effective in modeling crop phenology by aligning time series data with seasonal reference curves [40,41,42]. However, both methods assume symmetric treatment of deviations—i.e., upward and downward differences from the reference curve are penalized equally. This symmetry limits their ability to detect abnormal agricultural patterns such as abandoned planting or harvesting, which often exhibit directional NDVI deviations.

In practice, abandoned planting yields persistently low NDVI, while abandoned harvesting shows prolonged high NDVI post-maturity. In such cases, treating upward and downward deviations symmetrically—as DTW and TWDTW do—can lead to semantic misinterpretation. Two time series with drastically different agricultural meanings may produce similar alignment costs, reducing the ability to distinguish truly abnormal field conditions. Figure 6 illustrates these contrasting scenarios, where (a) shows undergrowth and overgrowth penalties associated with planting and harvesting failures, and (b) visualizes the post-maturity NDVI divergence caused by abandoned harvests.

To overcome this limitation, we propose the Bias-weighted Time-Weighted Dynamic Time Warping (BTWDTW) algorithm, which builds upon TWDTW by introducing an asymmetric residual penalty function. This function dynamically adjusts the alignment cost based on both the direction and magnitude of residual NDVI differences between the reference and observed time series. In doing so, BTWDTW places greater penalties on deviations that are agriculturally more indicative of failure (e.g., overgrowth following maturity or persistent undergrowth), improving the method’s ability to distinguish between sowing and harvest abandonment.

In TWDTW, time series $A=\{a_1,\dots ,a_n\}$ and $B=\{b_1,\dots ,b_m\}$ are aligned using a base distance matrix $S_{ij}={(a_i-b_j)}^2$, modulated by a temporal weight function. The time weight $w_{ij}$ penalizes matches between temporally distant points and is defined using a logistic function:

$$w_{ij}={\displaystyle \frac{1}{1+\mathit{exp}\left[-g\left(\left|i-j\right|-m_c\right)\right]}}$$

(3)

where $\mathrm{g}$ is the gain factor that controls the steepness of the penalty curve, and $m_c$ is the midpoint of the time axis. This weighting helps constrain the warping path to temporally plausible alignments, thereby improving phenological consistency.

BTWDTW extends this by integrating a directional bias penalty function into the cost calculation, enabling asymmetric treatment of residuals based on their sign and magnitude. The final distance metric is defined as follows:

$$D\left(i,j\right)=dist(x_i-y_j) w\left(i,j\right)\cdot b\left(x_i,y_j\right)$$

(4)

where $dist(x_i-y_j)$ is the pointwise distance, $w\left(i,j\right)$ is the standard time-weighting function from TWDTW, and $b\left(x_i,y_j\right)$ is a directional penalty function, defined as follows:

$$b\left(x_i,y_j\right)=\left\{\begin{array}{c}\alpha +\gamma \cdot \left({x_i-y}_j\right), if x_i>y_j\\ \beta +\gamma \cdot \left(y_j-x_i\right), otherwise\end{array}\right.$$

(5)

This function penalizes positive residuals (overgrowth) and negative residuals (underdevelopment) differently, enabling sensitivity to specific abnormal states. For abandoned planting (undergrowth), set $\alpha =0.5$, $\beta =1.5$, $\gamma =0.5$. For abandoned harvesting (overgrowth), set $\alpha =1.5$, $\beta =0.5$, $\gamma =0.5$.

3.5. Abandoned Cropland Mapping

NDVI phenology varies significantly across crops—for instance, wheat peaks earlier and declines rapidly, while spring crops like maize and sunflower peak later and sustain high NDVI longer. These differences make fixed thresholds unsuitable for detecting abandonment across all crop types.

To address this, we used a data-driven approach based on fields consistently planted with the same crop from 2019 to 2021. Crop-specific reference NDVI curves were generated, and BTWDTW similarity distances were calculated between each field and its corresponding reference. Thresholds for detecting significant deviation were defined as follows:

$$threshold=k\times \left(mean+std\right)$$

(6)

The coefficient k was optimized through comprehensive sensitivity analysis using our validation dataset. As demonstrated in Supplemental Section S1 and Table S1, k = 1.5 achieved optimal balance between detection completeness and reliability while maximizing overall accuracy and kappa coefficient.

Each field’s crop type was inferred from historical planting patterns; for rotational fields, crop type was predicted from the sequence or all reference curves were tested.

A two-stage classification was applied. If the growing-season distance exceeded the sowing abandonment threshold, the field was classified as unsowed. For remaining fields, harvesting-period distance was assessed; if it exceeded the threshold, the field was labeled unharvested. This method improves crop-specific detection accuracy by accounting for phenological variation and temporal deviations (Figure 7). The final classification outputs a spatial map categorizing each field into one of three types: unsowed, unharvested, or normally cultivated.

3.6. Accuracy Assessment

To quantitatively evaluate the effectiveness of the proposed method and verify the reliability of the abandoned cropland mapping results, we conducted a comprehensive accuracy assessment using visual validation samples. Specifically, a total of 2845 abandoned cropland sample points were compiled, yielding 8535 annual field records after merging across the three years. These samples served as reference data for validating classification outcomes.

Sentinel-2 data, with its high spatial resolution (10 m), was selected as the reference for this assessment due to its superior visual accuracy and spatial details. The high spatial resolution of Sentinel-2 allowed for more precise identification of abandoned croplands, providing detailed validation of the classification results. Furthermore, the frequent revisit cycle of Sentinel-2 data offered up-to-date imagery, which enhanced the temporal precision of the validation process, complementing the MODIS data and improving the overall reliability of the assessment. The classification performance of the proposed Bias-weighted Time-Weighted Dynamic Time Warping (BTWDTW) algorithm was systematically compared against the conventional TWDTW method. Both methods were applied to the same dataset to identify abandoned croplands, and their outputs were assessed using four standard accuracy metrics: user accuracy (UA), producer accuracy (PA), overall accuracy (OA), and the kappa coefficient.

4. Results

4.1. Pre-Conflict Crop Distribution

The ROCKET model was applied to classify crop types based on four vegetation indices (NDVI, EVI, LSWI, and NDYI) for each year from 2019 to 2021. The crop classification maps (Figure 8) reveal distinct spatial patterns of crop distribution across Ukraine from 2019–2021.

Statistically, the map revealed that wheat, sunflower, rapeseed, maize, soybean, and potato accounted for 30.59%, 12.69%, 2.49%, 51.46%, 1.12%, and 1.65% of the total crop area in Ukraine, respectively. Wheat is widely cultivated nationwide, with large contiguous areas concentrated in the eastern and southern regions, particularly in Donetsk, Zaporizhzhia, Kherson, Odesa, and Crimea. Maize is predominantly found in the central, northern, and western agricultural heartlands, including Kyiv, Zhytomyr, and Ternopil, where soil fertility and irrigation conditions are favorable. Scattered maize fields are also observed in Luhansk and neighboring areas in the east. Sunflower shows a more centralized distribution in the central–southern oblasts such as Dnipropetrovsk and Mykolaiv, benefiting from dry climates and abundant sunlight. Rapeseed is more dispersed, mainly across the central and southern regions, often overlapping with wheat and sunflower in crop rotation systems. Soybean and potato exhibit the most fragmented pattern, primarily located in the north and northwest, where it remains locally significant despite limited cultivation area.

The overall accuracy (OA), kappa coefficient, producer’s accuracy (PA), and user’s accuracy (UA) for each crop category are summarized in

Table 1. Accuracy results of crop type classification in Ukraine from 2019 to 2021.

Crop	PA	UA	OA	Kappa
Wheat	1.0000	0.9798	0.9567	0.9132
Sunflower	0.9663	0.9149
Rapeseed	0.9545	1.0000
Maize	0.9565	0.9429
Soybean	0.7797	0.9787
Potato	0.8868	0.9592

. The classification achieved an OA of 0.9567 and a kappa coefficient of 0.9132, indicating strong performance. PA values for most crops were high, with wheat reaching 1.0000, while sunflower and maize had PA values of 0.9663 and 0.9565, respectively. UA values were also high, with rapeseed achieving 1.0000 and soybean 0.9787, demonstrating the method’s effectiveness in accurately classifying different crops.

4.2. Post-Conflict Abandoned Cropland Distribution

Figure 9 illustrates the annual extraction results of abandoned cropland across Ukraine from 2022 to 2024. In 2022 (Figure 9a), abandonment was concentrated in southern oblasts such as Kherson, Zaporizhzhia, and Mykolaiv, closely aligned with early conflict zones. By 2023 (Figure 9b), the extent of abandonment decreased, but persistent patches remained in the south. In 2024 (Figure 9c), abandonment intensified and expanded northeastward, covering large parts of Donetsk, Luhansk, and Crimea. A zoomed-in view (Figure 9d) reveals clustered abandonment near major frontlines, providing finer-scale detail. To assess classification reliability, time series satellite imagery (Figure 9e) demonstrates a clear transition from active cultivation in 2022 to widespread unsowed and unharvested land in 2024. The red overlays confirm that abandoned parcels identified by the model correspond well with visual evidence, particularly in central pivot irrigation areas.

Overall, the BTWDTW-based method achieved higher classification accuracy than the TWDTW approach and the more conventional DTW approach. As shown in

Table 2. Accuracy results of sowing abandonment cropland detection in Ukraine from 2022 to 2024.

Method	Year	Unsowed		Sowed		OA	Kappa
		PA	UA	PA	UA
DTW	2022	0.7214	0.8423	0.7621	0.7519	0.7498	0.7402
	2023	0.7356	0.8502	0.7984	0.7785	0.7723	0.7841
	2024	0.7872	0.8438	0.7689	0.7942	0.7885	0.7698
	Total	0.7481	0.8454	0.7765	0.7749	0.7702	0.7647
TWDTW	2022	0.8200	0.9101	0.8403	0.8307	0.8398	0.8405
	2023	0.8237	0.9185	0.8784	0.8579	0.8520	0.8647
	2024	0.8671	0.9119	0.8476	0.8740	0.8681	0.8492
	Total	0.8369	0.9135	0.8554	0.8542	0.8533	0.8515
BTWDTW	2022	0.8003	0.9407	0.9635	0.8323	0.8852	0.8516
	2023	0.8405	0.9201	0.8912	0.8793	0.8624	0.8687
	2024	0.9334	0.9291	0.8603	0.9048	0.9201	0.9133
	Total	0.8581	0.9300	0.9050	0.8721	0.8892	0.8779

, the three-year average OA improved from 0.8533 to 0.8892, and the kappa coefficient increased from 0.8515 to 0.8779, indicating stronger model robustness and consistency. When compared with the traditional DTW method, which achieved an OA of 0.7702, BTWDTW outperformed it by a significant margin, demonstrating a more reliable and accurate detection of abandoned croplands.

Notably, for the sowing abandonment class, the average PA rose to 0.8581 and UA to 0.9300, outperforming TWDTW by 2.12 and 1.65 percentage points, respectively, and DTW by 11 and 8.46 percentage points, respectively. This suggests that BTWDTW not only enhances the detection of truly abandoned parcels but also reduces false positives. BTWDTW also improved accuracy for non-abandoned fields. The average PA increased from 0.8554 to 0.9050, and the UA from 0.8542 to 0.8721. These gains demonstrate the model’s enhanced sensitivity to NDVI curve deviations through directional weighting, which is not as effectively addressed by DTW and TWDTW.

Among all the years, the improvement in 2024 was most pronounced. In that year, the PA for abandoned cropland increased from 0.8671 (TWDTW) to 0.9334 (BTWDTW), while the kappa coefficient rose by 6.41 percentage points to 0.9133. This can be attributed to the more spatially aggregated distribution of abandoned land in 2024, which facilitated better detection performance.

4.3. Spatiotemporal Distribution of Abandoned Cropland

Between 2022 and 2024, Ukraine’s abandoned cropland area showed a distinct V-shaped trajectory (Figure 10), reflecting the evolving impacts of war on agricultural activity. The total abandoned area reached 28,184.20 km² in 2022 (7.21% of total cropland), declined to 24,007.65 km² in 2023 (6.14%), and peaked again at 33,278.34 km² in 2024 (8.52%).

Disaggregated by type, both sowing abandonment and harvest abandonment (unharvested cropland) reached high levels in 2022, measuring 11,641.57 km² (2.98%) and 16,542.63 km² (4.23%), respectively. This pattern reflected widespread systemic collapse during the initial stage of the conflict. In 2023, sowing abandonment dropped sharply to 3838.19 km² (0.98%), driven by the partial recovery of sowing activities in relatively secure areas. However, harvest abandonment rose to 20,169.46 km² (5.16%), likely due to logistical breakdowns during the harvest season and intensified attacks on agricultural infrastructure. By 2024, sowing abandonment rebounded to 17,269.45 km² (4.42%) due to renewed conflict along frontlines and disrupted spring operations. Harvest abandonment decreased slightly to 16,008.89 km² (4.10%).

Figure 11 presents kernel density estimations for sowing and harvest abandonment during the same period. The top row (Figure 11a–c) shows the spatial clustering of sowing abandonment. In 2022, hotspots were scattered across southern Ukraine, and became more fragmented in 2023. However, by 2024, high-density clusters emerged prominently in eastern oblasts, particularly Donetsk and Zaporizhzhia. In contrast, the bottom row (Figure 11d–f) displays harvest abandonment, which remained widespread across central and western Ukraine throughout all three years. Notably, in 2023, harvest abandonment temporarily expanded eastward, reaching parts of Luhansk, Donetsk, and Zaporizhzhia, before retreating in 2024 back toward the central and western heartlands. While sowing abandonment patterns exhibit strong spatial correlation with conflict zones, harvest abandonment appears more diffuse and persistent in less directly affected areas.

4.4. Abandonment by Crop Type

Abandonment patterns varied significantly across crop types, shaped by agronomic characteristics, logistical dependencies, and war-related disruptions. As shown in Figure 12, maize was the most severely affected crop, with an average total abandonment of 14,280 km² between 2022 and 2024, accounting for 53.6% of all abandoned cropland. Notably, over 66% of this abandonment was due to unharvested fields, indicating maize’s high vulnerability to transportation breakdowns, fuel shortages, and mechanical harvest failure.

Wheat ranked second in total abandonment (10.8%), with a more balanced structure: unsowed fields were dominant in 2022, while unharvested wheat surged in 2023 due to port blockades and depressed export prices. Sunflower and rapeseed experienced moderate levels of abandonment, mainly in the sowing stage, reflecting high input requirements and lower planting flexibility under instability. In both crops, over two-thirds of the abandoned area resulted from sowing failure rather than harvest disruptions. By contrast, soybean and potato had the lowest abandonment rates—below 1% of the national total—due to limited cultivation scale and concentration in relatively secure regions. Their abandonment was mainly unsown-based and remained consistently low across all years.

Overall, sowing and harvest abandonment exhibit distinct structural characteristics: sowing abandonment is composed of multiple crop types and broadly mirrors Ukraine’s overall planting structure, indicating widespread disruption during the sowing stage. In contrast, harvest abandonment is overwhelmingly dominated by maize.

5. Discussion

5.1. Conflict Zones and Spatial Correlation with Cropland Abandonment

The spatial correlation between cropland abandonment and conflict zones has been extensively studied, with numerous studies indicating a strong association between violent conflicts and reductions in both the total land area cultivated and the area harvested [11,43,44]. We employ abandoned cropland extracting results and the Armed Conflict Location and Events Data (ACLED) [45] to assess the impact of violent conflicts on cropland left unplanted and unharvested. As ACLED may have reporting lags, the most recent available version at the time of manuscript preparation (April 2025) was used in this analysis. From 2022 to 2024, a total of 140,163 conflict-related incidents were recorded in Ukraine, including 33,616 battles, 24,756 air/drone strikes, and 1112 instances of violence against civilians. These events have had a significant impact on agricultural activity and land abandonment patterns throughout the conflict period [46].

Figure 13a shows that unsowed cropland is highly concentrated in the eastern and southern frontline oblasts, with Donetsk, Luhansk, Kherson, Zaporizhzhia, and Crimea each showing over 10% of their cropland left unsown in 2024. This spatial pattern aligns with the intensity of conflict activity mapped in Figure 13b, which highlights these oblasts as hotspots of military operations. As detailed in Table S2, in 2022, the unsowed proportions reached 10.02% in Kherson, 6.22% in Zaporizhzhia, and 9.12% in Crimea. These percentages increased significantly over time: by 2024, Kherson and Crimea reported 22.23% and 22.66% unsowed land, respectively, while Zaporizhzhia reached 12.69%.

The spatial distribution of sowing abandonment became increasingly concentrated in high-conflict areas over time. Specifically, the share of sowing abandonment within these conflict-prone zones rose from 54.26% in 2022 to 68.25% in 2023, and reached 76.64% in 2024. This sharp rise reflects the escalating disruption of early-season agricultural activities due to prolonged warfare and territorial occupation [8,15]. Figure 14 further confirms this trend: the absolute unsowed area in Kherson and Zaporizhzhia exceeded 3000 km² in 2024, more than doubling their 2022 levels. These findings underscore the cumulative and compounding effects of sustained conflict exposure on early-season agricultural operations [27,43].

In contrast, harvest abandonment shows a markedly different spatial distribution. Areas such as Vinnytsia, Mykolaiv, and Cherkasy reported relatively high levels of unharvested cropland, even though they are located away from the immediate frontlines. Table S2 in the Supplementary Materials reveals that Vinnytsia had 6.35% of cropland unharvested in 2022, rising to 7.44% in 2023, while Mykolaiv consistently showed above 2% across the three years. This pattern suggests that harvest abandonment is more closely related to post-planting disruptions, such as logistical failures, labor shortages, and collapsing local markets, rather than direct combat activity. These regions may have managed to sow crops successfully but were unable to complete the harvest due to insecurity-related supply chain breakdowns and resource unavailability [11,30,43].

5.2. Shifts in Abandonment Patterns

The abandonment patterns observed from 2022 to 2024 (Figure 10) demonstrate how agricultural decisions were shaped by the evolving war environment in Ukraine. A significant shift occurred in 2023, with a marked decrease in unsowed areas and an increase in unharvested land. This trend reflects a temporary recovery in spring planting, driven by improved control over liberated areas and targeted support policies. However, this was followed by disruptions in the harvest season, caused by renewed conflict intensity and infrastructure collapse.

According to USDA statistics, Ukraine’s harvested area for major crops (wheat, maize, and oilseeds) in the 2023/24 season declined by 20–30% compared to pre-war levels [47]. Satellite data indicate that approximately 6% of wheat and 9% of sown maize were left unharvested by the end of 2023, with only 91% of maize harvested despite planting nearly 4 million hectares [26,48].

In 2023, Ukrainian forces regained partial control over agricultural regions such as northern Kharkiv and western Kherson, increasing the cultivable area. Government subsidies, input support, and low-interest loans—backed by international aid—enabled farmers to resume spring sowing on many lands. According to the Food and Agriculture Organization (FAO), emergency support packages, including seed distribution, fertilizer provision, and financial aid, played a critical role in stabilizing production and restoring farming activities in reclaimed areas [49,50]. Consequently, the national unsowed rate dropped markedly compared to 2022.

However, the harvest season was severely disrupted. The destruction of the Kakhovka dam caused flooding and long-term irrigation failure across southern oblasts. Meanwhile, military operations in agricultural zones intensified, including shelling, field burning, and aerial attacks, resulting in further damage to cropland and infrastructure. Logistical constraints and security threats affected both Ukrainian-controlled and Russian-occupied territories, complicating harvesting and transportation efforts [51].

This V-shaped trend—initial collapse in 2022, partial recovery in 2023, and renewed disruptions thereafter—is distinct from abandonment trajectories observed in other conflict-affected countries. For example, in South Sudan, cropland abandonment between 2016 and 2018 followed a more persistent decline due to prolonged insecurity, displacement, and lack of coordinated recovery mechanisms [11]. In Syria, the civil war disrupted land use primarily through long-term occupation and regime changes, resulting in stable but suppressed cultivation levels over several years without significant seasonal rebounds [10]. In contrast, Ukraine’s temporary resurgence in 2023 was closely tied to targeted policy interventions and international aid, underscoring the role of external support in moderating war-induced agricultural collapse.

5.3. Crop-Specific Abandonment Patterns

Violent conflicts are strongly associated with significant changes in cropping patterns: farmers tend to reduce the land area allocated to long-term crops, while increasing the share of land allocated to short-term crops like cereals and seasonal crops. As illustrated in Figure 15, the cropping calendar for Ukraine reveals the seasonal rhythms and critical periods for various crops, which help contextualize these abandonment patterns. Abandonment patterns varied significantly by crop type in Ukraine (Figure 12). Maize, which is typically sown in the spring and harvested late in the season, exhibited the highest levels of unharvested abandonment. The vulnerability of maize to unharvested loss is tied to its longer maturation cycle, heavy machinery dependence, and reliance on functioning post-harvest logistics such as storage, drying, and transportation. When frontline disruptions occurred mid-year, maize fields were often left unharvested, as the logistical infrastructure became insufficient to support timely harvest operations. This relationship is summarized in Figure 16, which outlines the key factors contributing to maize’s vulnerability to harvest abandonment during conflict.

In contrast, wheat, sown in autumn and harvested earlier than other crops, showed higher rates of sowing abandonment. Oil crops like sunflower and rapeseed experienced higher sowing abandonment, likely due to delayed fertilizer and seed availability, compounded by disrupted market flows. Soybean and potato fields, mainly in western oblasts, had the lowest abandonment rates, likely due to localized control, smaller farm sizes, and better manual management. These areas saw fewer disruptions, allowing for more stable agricultural activities. This differential abandonment pattern underscores the importance of understanding crop-specific vulnerabilities in the context of war and how crops’ physical and economic characteristics interact with conflict dynamics.

5.4. Limitations and Future Work

The proposed method relies on dense time series satellite observations to capture crop growth dynamics. However, the spatial resolution of MODIS data, while suitable for large-scale mechanized farming regions like Ukraine, limits its applicability in heterogeneous agricultural landscapes with smaller field sizes. To enhance spatial accuracy and generalizability, future research could integrate higher-resolution datasets such as Sentinel-2 and synthetic aperture radar (SAR) imagery. Data fusion [52,53] techniques can improve the method’s flexibility across different agro-ecological zones, allowing for a more accurate identification of abandonment at the parcel level. Furthermore, approaches such as mixed pixel decomposition [54] and super-resolution reconstruction [55] offer the potential to refine spatial resolution, thereby increasing the method’s applicability at finer spatial scales.

Second, this study focuses primarily on cropland abandonment between 2022 and 2024, without fully capturing the complexity of Ukraine’s multi-year crop rotation systems. In many regions, land use varies cyclically across years, and distinguishing between rotational fallow and true abandonment remains a challenge [56]. Long-term monitoring using extended time series data would allow for more accurate classification of persistent versus temporary abandonment. Future work could also explore post-conflict recovery trajectories—evaluating which areas resume cultivation through policy interventions, and which may face permanent agricultural loss due to depopulation or environmental degradation [25].

6. Conclusions

This study proposed and implemented a novel framework for cropland abandonment detection, integrating MODIS time series data with a Bias-weighted Time-Weighted Dynamic Time Warping (BTWDTW) method. This framework enables a fine-grained distinction between unsowed and unharvested cropland, addressing key methodological limitations in previous conflict-related agricultural studies.

The results reveal a V-shaped trajectory of abandonment in Ukraine during 2022–2024, with total abandoned areas estimated at 28,184 km² in 2022 (7.21% of total cropland), decreasing to 24,008 km² in 2023 (6.14%), and then peaking at 33,278 km² (8.52%) in 2024. Spatial analysis indicates that sowing abandonment exhibited strong clustering in eastern frontline oblasts—Donetsk, Luhansk, Zaporizhzhia, and Kherson—where over 70% of unsowed areas overlapped with conflict zones. In contrast, unharvested abandonment followed more dispersed patterns.

The study also revealed significant crop-specific differences. Maize accounted for over 53% of total abandonment, with more than two-thirds of it unharvested, while wheat contributed 10.8%, with a more balanced structure between unsowed and unharvested loss. Sunflower and rapeseed showed moderate sowing abandonment, whereas soybean and potato, concentrated in safer western oblasts, experienced minimal losses.

Overall, the proposed framework achieved an average accuracy of 88.9% and kappa coefficient of 0.878, demonstrating robust performance across years and categories. These findings contribute valuable insights into the spatial, temporal, and structural dynamics of war-induced agricultural disruption in Ukraine and offer a scientific foundation for post-conflict recovery and food security planning.