Monitoring winter crop areas during wartime: remote sensing support for Ukraine’s agricultural statistics

Russia’s full-scale invasion of Ukraine in February 2022 disrupted grain markets and agricultural statistics alike. The war triggered sharp commodity price fluctuations¹ and logistical bottlenecks, while Ukraine’s agricultural exports, though partially sustained through the Black Sea Grain Deal and later via European “Solidarity Lanes” (European Council, 2023; European Commission, 2025), continued under highly uncertain conditions. As one of the world’s major breadbaskets and a key wheat supplier to the Middle East and North Africa, Ukraine plays a crucial role in global food security (European Council, 2025). Amid these shocks, reliable information on Ukraine’s crop production became increasingly critical for maintaining market transparency and global food security².

Reports such as the United States Department of Agriculture World Agricultural Supply and Demand Estimates (USDA WASDE) and Joint Research Centre Monitoring Agricultural ResourceS (JRC MARS) bulletins publish production estimates to increase market transparency and actively inform trade decisions^2,3,4. Before the war, large-scale commercial farms, which produced about 85% of Ukraine’s wheat, were required to report planted and harvested areas, as well as production, to the State Statistics Service of Ukraine (SSSU). The remaining 15% was estimated from sample surveys of small producers (SSSU, 2024). However, from March 3^rd, 2022 to August 18^th, 2025 Law №2115-IX suspended mandatory reporting for these agricultural statistics due to martial law (Government of Ukraine, 2022, Government of Ukraine, 2025). While many farmers continued to provide data voluntarily, surveys of smallholder producers were halted for security reasons, creating national-level gaps in official statistics for Government-held areas. Data flow from Russian-controlled territories ceased abruptly from the onset of the war to the present. In 2023, 2024, and 2025, USDA stopped reporting on Russian-controlled territories except for Crimea, where agro-statistics are sourced from Rosstat, the Russian statistical agency (USDA, 2022; 2023; 2024; 2025). The statistical data blackout in Russian-held territories required alternative, remotely actionable solutions for compensating the loss of reporting-based agricultural production estimates.

Several studies used remote sensing data to estimate crop production losses in Ukraine. Dai et al. ⁵ mapped grid-level crop abandonment rates, then proportionally adjusted pre-war crop production numbers based on abandoned area rates. He et al. ⁶ and Chen et al. ⁷ first assessed changes between pre- and during the war vegetation index dynamics and then applied change ratios to estimate lost agricultural productivity over abandoned and non-abandoned croplands in Ukraine. Jia et al. ⁸ used remote sensing data to map winter cereals and rapeseed and to estimate corresponding yields. They subsequently estimated production for season 2022 and assessed losses by comparing to pre-war levels. While estimating these losses provided valuable insights, official agencies had an even more critical need for timely information on ongoing agricultural production to contextualize these losses and support food security assessments, policy planning, and recovery efforts.

Since the beginning of the war, only Qadir et al. ⁹, Qadir et al. ¹⁰ and Qadir et al. ¹¹ developed an operational, statistically sound, remote sensing-based only framework for generating sunflower planted area estimates in war-torn Ukraine. In this paper, we address the complementary need for an alternative source of reliable winter wheat and rapeseed planted area statistics, given Ukraine’s global importance in the production and export of these crops. Our objectives are twofold: (i) to support and strengthen Governmental statistics, which are inferred from incomplete datasets as compared to pre-war, and to (ii) track planted areas in support of production estimates in Russian-held territories, from where data flow ceased entirely.

Quantifying the uncertainty of area estimates is extremely important to official statistics. For instance, the USDA NASS June Agricultural survey aims for a coefficient of variation lower than 2% for maize, soybean, and winter wheat area estimates¹². Regardless, most remote sensing-based area estimates in literature do not quantify uncertainty and often use the pixel-counting estimator for area estimation: the multiplication of pixel area by the number of pixels in a map class^5,6,8,13,14. Such an estimator is biased because of omission and commission errors in the classification maps and mixed pixels at class boundaries^15,16,17,18 and is discouraged for estimating crop area. Instead, maps can be used as a stratification tool in the sampling design to further employ estimators that are unbiased^15,16,17,18. Studies by Song et al. ¹⁹, Li et al. ²⁰ and King et al. ²¹ established the feasibility of planted area estimation using a two-stage sampling approach for maize and soybean area estimation in the USA, Argentina, and China. Ideally, once drawn, each sample unit should be visited on the ground to collect the “true” crop type label. However, when ground visits may jeopardize lives^22,23, remote sensing imagery interpretation is a safe and cost-efficient alternative for reference data collection so long as it is carried out by experts with a deep and comprehensive understanding of the agricultural landscape. Olsen et al. ²⁴, Kerner et al. ²⁵, Li et al. ²⁶ and Skakun et al. ²⁷ substituted field visits with remote sensing imagery interpretation to collect reference data. They estimated active cropland areas and crop loss based on satellite-based classification maps and stratified random sampling in conflict-affected Sudan, Tigray, Syria, and Ukraine regions.

But it is worth emphasizing that, unlike ground surveys, which provide a gold-standard dataset, remote sensing imagery interpretation can suffer from interpreter variability, potentially affecting area estimates and their precision²⁸. Annotation quality ultimately depends on the annotator’s domain knowledge, which, when insufficient, can lead to severely flawed labeling and estimates. For instance, Chen et al. ⁷ overestimated Ukraine’s winter wheat, rapeseed, sunflower and maize planted areas in 2020 by respectively 35%, 209%, 14% and 72% relative to official statistics (before the war): they communicated “unbiased estimates” (± standard errors) of 8.7 ± 0.4, 3.4 ± 0.4, 7.3 ± 0.5 and 8.9 ± 0.5 million hectares (Mha) for wheat, rapeseed, sunflower and maize, while the State Statistics Service of Ukraine reported 6.5, 1.1, 6.4 and 5.2 Mha respectively for these crops, in 2020 (SSSU, 2024). In every paper relying on annotated data for validation, clear guidelines on how reference data is interpreted from remote sensing imagery should be provided.

In this paper, we leveraged the specific phenological and temporal traits of winter crops²⁹ and rapeseeds^30,31 to operationally generate wall-to-wall, unsupervised, and in-season crop type maps³², based on Planet satellite data, for each agricultural season from 2022 to 2025. Maps were then used for stratification in a stratified random sampling design for winter crop planted area estimation, separately in Government-controlled and Russian-held territories. After annotating sample units, we estimated areas and uncertainties using a stratified estimator, which is unbiased³³. Planted area estimates were compared against SSSU agricultural statistics in Government-controlled territories, reinforcing confidence in numbers for Russian-controlled areas. In the absence of feasible ground data collection, we established reference data through systematic remote sensing imagery annotation. Given past issues with annotation reliability⁷ and often lacking explanations on photointerpretation with remote sensing imagery^{13,24,34,35,36,37}, we also present a set of guidelines for visually annotating winter cereals, rapeseed, spring and summer crops, and non-cultivated cropland classes in Ukraine to support consistency and accuracy in future assessments. During a time of war, this work provided the only in-season planted areas derived from an unbiased estimator, for winter wheat and rapeseed across all of Ukraine, one of the world’s largest exporters of both commodities (USDA, 2022). This work is part of a collaboration between NASA Harvest (NASA Harvest), the State Statistics Service of Ukraine (SSSU), the Ministry of Agrarian Policy and Food of Ukraine (MAPFU), the Food and Agriculture Organization (FAO), and the European Development and Reconstruction Bank (EBRD). It aimed at operationally generating crop areas, yields, and production estimates, based on remote sensing data, in war-affected Ukraine (FAO Investment Centre).

Winter crop planted area estimates

Since the onset of the full-scale war, we seasonally estimated planted areas and associated standard errors for winter wheat, other winter cereals, and rapeseed, by military control status in Ukraine. We used military control lines as of July 11^th, 2022, July 31^st, 2023, May 22^nd, 2024, and June 17^th, 2025 in final area estimates delivered to the SSSU.

In Government-controlled territories (Fig. 1), winter wheat planted area decreased by 9% from 2022 to 2023, from 5.14 ± 0.52 Mha to 4.68 ± 0.25 Mha (Here and throughout the paper, uncertainties are represented as one standard error). Wheat planted areas slightly increased in 2024 compared to 2023, with 4.73 ± 0.28 Mha planted, and remained nearly stable in 2025, with 4.84 ± 0.25 Mha planted. In 2022, we recomputed planted area estimates based on the September 26th, 2022, military control line, after significant portions of Sumy, Kharkiv, and Mykolaiv oblasts were reclaimed by Ukrainian forces. Updated estimates showed planted areas of 5.39 ± 0.56 Mha, 0.9 ± 0.09 Mha, and 1.18 ± 0.17 Mha for winter wheat, other winter cereals, and rapeseed, respectively. During the same period, rapeseed planted area increased from 1.16 ± 0.17 Mha in 2022 to 1.96 ± 0.21 Mha in 2023. Following this increase, rapeseed planting returned to a level comparable to 2022, with 1.29 ± 0.15 Mha planted in 2024. In 2025, the rapeseed area increased again to 1.47 ± 0.23 Mha.

In Russian-controlled territories as of the July 2022 military control line, winter wheat planted area decreased by 10% from 2022 to 2023, dropping from 2.06 ± 0.16 to 1.86 ± 0.09 million hectares (Mha) (Fig. 2). In our September 2022 update, following the reclamation of land by Ukrainian forces, we estimated that 1.80 ± 0.16 Mha were planted to winter wheat, 0.3 ± 0.03 Mha to other winter cereals, and 0.16 ± 0.01 Mha to rapeseed. In 2024, significant disruption in planting led to a one-third decline in winter wheat area compared to 2023, with only 1.39 ± 0.08 Mha planted. Over the same period, rapeseed area fell from 0.17 ± 0.01 Mha in 2022 to 0.14 ± 0.03 Mha in 2023. Planting was severely disrupted in 2024, with rapeseed area declining fourfold to just 0.03 ± 0.01 Mha. In 2025, winter wheat and rapeseed planted areas saw modest recoveries to 1.59 ± 0.1 Mha and to 0.05 ± 0.01 Mha.

Winter crop maps and accuracy estimates

Trends observed in planted area estimates are also reflected in the maps (Fig. 3). The overall distribution of winter crops remained stable over the three years, except for the absence of cultivation along the front. Notably, a larger area of rapeseed was planted in 2023 across Government-controlled territories, but was negligible in Russian-held areas.

Detailed yearly confusion matrices and accuracy estimates (area-based using unbiased estimators) can be found in Tables S1 to S4. Across seasons and control zones, winter cereals (including winter wheat) consistently showed higher producer’s accuracy (PA, average = 87.6%) than user’s accuracy (UA, average = 75.3%). UA corresponds to the proportion of elements predicted to class i that rightfully belong to i. PA corresponds to the proportion of elements that truly belong to class i that were mapped to i. Most commission errors (1-UA) arose from confusion with summer crops in 2022–2023 or non-winter crops in 2024–2025, while omission errors (1-PA) reflected limitations of the ESA WorldCereals cropland mask^38,39, which excluded some true winter crops as evidenced during reference data collection.

In contrast, rapeseed classification was less stable and more sensitive to seasonal and sensor differences. Averaged across all years and control zones, UA was 82% and PA was 72%, with UA typically exceeding PA except in Russian-controlled areas in 2022 and 2025. Omission errors mainly involved confusion with non-crop, summer crop, or winter cereal classes, again linked to the cropland mask. From 2022–2024, reliance on Planet two-week composites introduced temporal inconsistencies that obscured flowering signals. In 2025, use of Sentinel-1 VH polarization composites produced asymmetric outcomes: in Government-controlled territories, UA = 100% and PA = 55 ± 9%, while in Russian-held regions, PA = 100% and UA = 30 ± 7%. Visual inspection confirmed artifacts, especially in Crimea, indicating that while winter cereal mapping remained robust and consistent, rapeseed accuracy was more variable and sensor dependent.

When the Russian full-scale invasion of Ukraine began in February 2022, wheat prices spiked on international markets, triggered by the uncertainty around Ukraine’s production and export capacity to come. However, despite warnings of a “Coming food catastrophe” (The Economist, 2022), “Half of Harvests […] Wiped Out by War” (Bloomberg, 2022), “harvests could be halved […]” (The Guardian, 2022) relayed in media, our results demonstrated that total Ukraine’s wheat harvested area did not drop drastically in 2022 (Becker-Reshef et al., 2022), nor did planted areas in upcoming years. In this situation, timely tracking of crop production, starting with winter crop planted areas, helped keep the market transparent and ultimately participated in lifting initial export licensing on wheat by July 2022 (FAO, 2022). Our results showed that, when aggregated at Ukraine-scale, wheat areas approximately decreased by 3 to 15% (0.19 to 0.92 Mha) compared to the USDA 2017 to 2021 average (USDA PSD-Online). For rapeseed, a cash industrial crop, areas increased by 15 to 48% or between 0.19 and 0.97 Mha more planted, compared to the USDA 5-year pre-war average.

Between 2022 and 2025, NASA Harvest supported the SSSU by providing in-season estimates of winter wheat, rapeseed, and other winter cereals planted areas within Government-controlled territories. NASA Harvest’s remote-sensing-based estimates strengthened numbers reported by the SSSU across conflict years. Figure 4 compares NASA Harvest’s planted area estimates and the SSSU final reported areas for winter wheat, rapeseed, and other winter cereals. SSSU statistics reported here are final for the season 2022 to 2024 but may still be revised for season 2025 (report as of September 2025).

For winter wheat, the target commodity in this paper, SSSU’s reported areas, based on voluntary reporting, consistently fell within one standard error of NASA Harvest estimates across 2022 to 2025 seasons (Fig. 4a). The maximum difference between SSSU and NASA Harvest was 0.21 Mha (<5%) in 2023. We found the best alignment in 2024, with a difference of less than one percent, equal to 0.02 Mha. The RMSE of 0.138 Mha and a bias of −0.002 Mha indicate that NASA Harvest estimates closely align with SSSU data, with no evidence of systematic over- or underestimation.

As for winter rapeseed, SSSU’s reported areas for rapeseed were within the bounds of one standard error from the NASA Harvest estimates, in seasons 2022 and 2024. The 2023 NASA Harvest planted area estimate was significantly higher than SSSU’s reported area, at 1.96 ± 0.21 Mha compared to SSSU’s 1.39 Mha (Fig. 4b). Results from other data providers converge on estimating a higher rapeseed area than SSSU: USDA reported 1.54 Mha of winter rapeseed in Government-controlled territories (Table 1), and JRC reported a planted area of 1.7 Mha for full Ukraine in their June 2023 MARS bulletin⁴⁰. Similarly, in 2025, SSSU reports 1.18 Mha planted, to rapeseed while NASA Harvest, USDA (USDA PSD-Online) and JRC⁴¹ respectively estimated 1.47 ± 0.23 Mha, 1.27 Mha and 1.311 Mha. Due to those inconsistencies, the RMSE for rapeseed was 0.321 Mha, and the bias was 0.23 Mha, indicating slight misalignment and overestimation of planted areas by NASA Harvest, as compared to SSSU. Some limitations might come from the area estimation process. Small strata – like, in our case, rapeseed with a 3% stratum weight – may be affected by higher standard errors and increased risk of divergence in area estimation. As recommended by Olofsson et al. ³³, we systematically oversampled rapeseed to control these phenomena.

For other winter cereals, NASA Harvest estimates aligned well with SSSU’s in 2022, and exceeded SSSUs by less than 0.15 Mha between 2023 and 2025 (Fig. 4c). At the time of writing this paper, SSSU statistics that were made available for season 2025, only covered wheat, rapeseed and barley. The RMSE and bias are comparatively low, respectively 0.1 Mha and −0.055 Mha, indicating slight underestimation of other winter crop areas. The overall agreement between the two data sources, especially for winter wheat, reinforces the validity of our methodology and enables trust in area estimates for the Russian-controlled territories, where ground surveys and statistics are not available.

We further cross-compared our winter wheat and rapeseed planted area estimates with USDA’s harvested area reports from 2022 to 2025 (Table 1). USDA combines spring and winter planted areas into one single number, respectively, for wheat and rapeseed. To ensure comparability, we derived winter-to-spring area ratios from SSSU yearly statistics (SSSU, 2024) and applied these to USDA’s reported crop areas. USDA reported harvested areas for all of Ukraine in September 2022 (USDA, 2022). In 2023, 2024, and 2025, USDA discontinued their reporting on Russian-held territories: numbers include only Government-controlled territories plus Crimea, with Crimea’s data sourced from Rosstat, the Russian Statistical Agency (USDA, 2023; 2024; 2025).

In 2022, USDA’s harvested area estimates for winter wheat in full Ukraine were 1.82 Mha below NASA Harvest’s planted area estimates, aligning more closely with SSSU and NASA Harvest numbers for Government-controlled territories. We note that after 2022, USDA’s estimates were slightly higher than those from SSSU, but within one standard error of NASA Harvest’s Government-controlled estimates. For rapeseed, USDA and NASA Harvest estimates were closely aligned across seasons, reflecting that major rapeseed production areas are outside Russian-controlled territories. After 2022, only a few data providers, including JRC⁴² and NASA Harvest, continued monitoring Russian-held territories, demonstrating the importance of Remote Sensing analysis to maintain critical market information and transparency in times of war.

Estimating crop planted area using in-season maps for stratification within the sampling design typically requires a large number of sample units, representing significant investments in time and funding⁴³. To optimize ground data collection over extended regions, two-stage area frame and regression estimators are often employed, which involve visiting multiple spatially close points rather than randomly scattered ones^19,20,21. However, in our case, ground data collection was prevented by the war in Ukraine, particularly in Russian-controlled territories. We therefore relied on remote sensing imagery interpretation as an alternative, safe, and cost-effective way to collect reference data. Freed from constraints on the spatial distribution of validation points, we implemented a stratified random sampling approach, as outlined in Olofsson et al. ³³, to estimate planted areas in Government and Russian-controlled territories. Despite the availability of methods for generating in-season field boundaries in Ukraine⁴⁴, we preferred using map pixels as spatial assessment units. Using polygons as assessment units would have unnecessarily complexified the area estimation task⁴⁵ as field boundary delineations are imperfect⁴⁴, leading to within-polygon heterogeneity and possible omissions of croplands in the dataset. Thus, in our stratified random sampling design, map classes were the strata and map pixels were the sample units. In each stratum sample units were drawn as a simple random sample. Because certain parts of the non-cropland stratum from the map may have been misclassified and actually been cropped, we sampled this stratum in addition to the cultivated strata. We reduced the annotator’s bias²⁸ using clearly defined annotation guidelines, applied on multiple satellite image sources (two-week Planet composites, daily Planet imagery, and Sentinel-2) and vegetation index time-series. Additionally, two to three annotators labeled each sample unit, followed by a review of discrepant labels until a consensus was reached.

We acknowledge several limitations in using remote sensing for reference data collection. First, the approach is restricted to crop types or groups that can be reliably identified through visual inspection of satellite imagery time-series. Clear phenological traits, such as rapeseed’s yellow flowering^30,31, sunflower’s heliotropic behavior^9,10,11, or distinct sowing windows for winter and summer crops (USDA crop calendars), are required. While supervised models can separate summer crops using ground data, certain crop types like soybean and maize remain visually indistinguishable, preventing area estimates without field validation. Second, satellite-based annotation is also time-intensive. Based on our experience, one trained annotator can review ~400 sample units per full workday. Thus, annotating 1600 units with two people requires 2 × 4 workdays per estimation cycle for Ukraine. Despite the effort, this method remains more cost-efficient than field visits. Lastly, reliable annotation depends heavily on domain expertise. Even minor disagreements on the labels should be documented and discussed collectively to improve consistency and understanding. To support training of future annotators, we openly share (i) annotated sample units from 2022–2025, including locations, strata, and class labels, and (ii) a Sentinel-2–based Google Earth Engine tool used for annotation in our workflow (see Section Data Availability).

This work emphasizes the pivotal role that satellite-based estimates play in informing agricultural commodity markets and policy. We provide an important and missing piece of information to official statistics by keeping track of the planted area in Russian-held territories. It is also the cornerstone of a continued partnership with the SSSU, MAPFU, FAO, and EBRD, aimed at complementing traditional survey-based statistics with remote sensing-based production estimates in Government-controlled territories (FAO Investment Centre). This paper details the winter crop planted area estimation component of the initiative, which, together with yield estimates derived from the ARYA model^46,47, enabled calculation of winter wheat and rapeseed production across Ukraine. We conclude by citing Mr. Taras Vysotskyi, First Deputy Minister of Agrarian Policy and Food of Ukraine:

“The timeliness, objectivity, and transparency of the information remain essential factors in shaping a balanced and effective agrarian policy. We acknowledge and appreciate the vital role of NASA Harvest’s solutions and calculations in assessing the real state of agricultural production development in Ukraine, ensuring regional food security, and determining the export potential to support global food security.”

Workflow overview

Figure 5 outlines the workflow for sample-based crop planted area estimation, further detailed in Sections Satellite data to Remote Sensing imagery interpretation. First, remote sensing imagery from Planet, Sentinel-1, and Sentinel-2 is used to produce an in-season unsupervised map of winter crops (Section Domain knowledge for multi-year crop mapping). This map serves as a stratifier to draw two separate stratified random samples, one for each occupation status. Both samples are annotated through remote sensing imagery interpretation, following rigorous sample unit annotation guidelines. Finally, areas and map accuracies are estimated using the approach described by Olofsson et al. ³³, yielding in-season planted area estimates for Government-controlled and Russian-held territories (Section Sample-based crop area estimation and accuracy assessment).

Satellite data

For seasons 2022 to 2024, we used daily, 3-meter spatial resolution, four bands (blue (490 nm), green (565 nm), red (665 nm) and near infra-red (865 nm)) Planet data, composited into two-week periods from February 24^th, 2020 onwards, until May 27^th, 2024 (Planet, Planet Technical Specifications). Analysis-ready data were transferred from Planet compute environments to Google Cloud Storage and subsequently ingested into Google Earth Engine⁴⁸ for further processing. For each Planet two-week composite, we computed 14 vegetation indices described in Table S5.

For season 2025, we used Sentinel-2 imagery from the 1^st of November 2024 to the 31^st of May 2025. We masked out invalid pixels using the Scene Classification Layer (SCL). We then computed temporal features named on “S2_[…]” in Table S5.

Across seasons, to cope with cloud cover, we used Sentinel-1 synthetic aperture radar imagery ground range detected data, freely available in GEE. We created median VV, VH, and VV/VH backscatter composites for two-week time intervals, matching the Planet data compositing periods.

Ancillary data

We selected the ESA-World Cereals 2021 cropland product for this study. Classification accuracies over Europe are high, with a 96.6% user’s accuracy and a 92.0% producer’s accuracy for their active cropland layer^38,39. To account for the climatic gradient across the country, we adapted the USDA-FAS agro-ecological zoning (AEZ) (USDA AEZs) of Ukraine to align with oblast boundaries. The AEZs are highlighted in green in Figure S2.

To estimate unbiased planted area per military control status, we used boundaries provided by the Institute for the Study of War and AEI’s Critical Threats Project (ISW and AEI) (Red lines in Figure S1).

Domain knowledge for multi-year winter crop mapping

Drawing on crop calendar information (USDA crop calendars), and our extensive previous experience in crop monitoring in Ukraine^{27,29,46,49,50,51}, we developed an in-season, deductive, phenology-informed mapping approach that relies on different crop phenological characteristics:

In Ukraine, winter crops are typically sown from September until the end of October, followed by a green-up period in November. While some winter crops exhibit clear signals, detecting autumn green-up in remote sensing imagery is not always straightforward. Factors such as management practices⁵², soil conditions⁵³, cultivar⁵⁴, and weather conditions⁵⁵ can influence or delay crop establishment. Additionally, green-up detection using optical sensors can be obstructed by cloud cover⁵⁶ and snow⁵⁷. After undergoing vernalization⁵⁸, winter crops resume vegetative growth in early spring. The peak vegetative stage is typically reached around May, coinciding with approximately 1000 °C of accumulated growing degree days since January 1st⁴⁹, with crops generally ready for harvest by July. During the flowering phase, rapeseed can be differentiated from winter cereals due to its bright yellow flowers, which are oriented skyward, producing a strong yellow reflectance signal in optical satellite imagery^30,31. Summer crops green up starting from mid-May, peak in July–August and get harvest ready from mid-September onward (USDA crop calendars).

From 2022 to 2024, we leverage these crop-specific spectro-temporal signals as inputs into a K-means clusterer for mapping winter cereals, rapeseed, summer crops, and non-cultivated cropland along the season. For each mapped crop type and, to account for the climatic gradient in Ukraine, per AEZ, we train and infer a separate unsupervised classifier. We first compute 14 vegetation indices and three SAR features from Planet and Sentinel-1 composites over relevant in-season mapping periods (Table S5). For model training, we randomly generate several thousand samples within cropland and extract the corresponding feature values stacked into a single input vector for each sample. Next, we train a K-means classifier on sampled vegetation index values and SAR features and infer the model at the AEZ level. Finally, we assign each cluster to its corresponding crop type based on Planet two-week imagery composites, vegetation index time series, and precise assignment guidelines presented in the section Remote Sensing imagery interpretation, Figs. 6–11. The classification process starts from the ESA World Cereals cropland layer^38,39, within which winter crops are first identified and subsequently subdivided into winter cereals and rapeseed, by end of May. From the remaining cropland area not classified as winter crops, summer crops are then mapped. Any residual cropland that is not attributed to either winter or summer crops is labeled as non-cultivated for the season. Resulting maps are then used as stratifiers to guide our sampling frames and estimate unbiased areas.

In 2025, the mapping steps remained consistent compared to previous seasons. Expanding on work by Skakun et al.²⁹ we substituted K-means by a Gaussian Mixture Model trained on four temporal features extracted from Sentinel-2 Normalized Difference Vegetation Index time series and Sentinel-1 VH polarization time series, for mapping winter cereals and rapeseed.

Sample-based crop area estimation and accuracy assessment

Each year, we use our in-season crop type map for stratification to generate a stratified random sample for unbiased area estimation across three to five strata. While this paper focuses on winter crop mapping and area estimation, we also map summer crops and classify residual, non-planted, cropland as non-cultivated. We independently assess planted areas in the Government- and Russian-controlled territories.

Table 2 presents the sampling designs along with stratum weights (W_i) and number of sample units (n_i) for each military control status. The stratum weight is the biased area estimate, derived from the map through pixel counting. Eq. (13) in Olofsson et al.³³ can be used for estimating the number of sample units required when targeting overall accuracy as the estimation objective. Sample units should then be allocated proportionally according to stratum weight. However, when sample weights are small, proportional allocation may result in very low sample counts, unless the total sample number is very large. In such cases, Olofsson et al.³³ recommend slightly over-sampling small classes. While suboptimal sample allocation may be reflected in standard errors, it will not compromise the unbiasedness of the estimates.

In the 2022 season, we used the military control line from the 11^th of July to separate Government and Russian-controlled territories (Figure S1A) and estimated areas for four strata: winter cereals, rapeseed, summer crops, and non-crops. Non-cultivated fields were integrated into the non-crop stratum due to their low sample weight. In 2023, cropland abandonment became distinctly visible along the frontline, which we set to the 31^st of July (Figure S1B). As a result, we estimated unbiased areas for the non-cultivated cropland stratum, regardless of its low sample weights. In 2024 and 2025, we split the area estimation tasks for winter and summer crops. We finalized our planted area estimates for winter cereals, rapeseed, and non-winter crop classes based on the military control status from the 22^nd of May 2024 (Figure S1C) and 17^th of June 2025 (Figure S1D).

Planet 5*5-meter spatial resolution pixels were used as the spatial assessment unit, also called “sample units” in seasons 2022 to 2024. In season 2025, Sentinel-2 10*10-meter pixels were used as sample units. Two to three experienced labelers annotated all sample units to classes corresponding to the strata. Our primary data sources for sample unit annotation are Planet two-week image and NDVI time-series, alongside daily Planet images. Additionally, we used openly available Sentinel-2 optical imagery composited to two-week intervals matching Planet, especially in 2025. Two-week composite time series are generally sufficient to distinguish between winter, spring, and summer crops, non-crop, and non-cultivated cropland classes. However, short-timed events, such as rapeseed flowering could easily be missed on two-week compositing intervals. We thus systematically verify for a rapeseed flowering signal, on daily Planet imagery, all sample units annotated to winter crop. Detailed visual annotation guidelines are provided in Section Remote Sensing imagery interpretation. Shapefiles of sample unit locations, stratum and annotated class values are openly available for seasons 2022 to 2025 on GitHub (see Section Data Availability). Sampled points and class values can be used for training future annotators or supervised machine learning models. Codes to replicate our Sentinel-2 based annotation software on GEE are openly available on GitHub (see Section Data Availability).

Following mathematical formulae to calculate unbiased planted areas, users, producers and overall accuracies next to associated standard errors are cited from Olofsson et al. ³³. As per Eq. 1, we computed a confusion matrix based on mapped area proportions rather than sample counts. The proportion of area of map class i and of reference class j in the confusion matrix is noted p_ij. The sample-based estimator (hat{p}) ({hat{p}}_{{ij}}) of p_ij. (Eq. 1), is used to compute overall, producer’s and user’s accuracy per class (Eqs. 2 – 4).

where W_i is the proportion of area mapped as class i, n_ij is the number of sample units of map class i and of corresponding reference class j and n_i are total number of sample units mapped to class i.

Overall accuracy (OA) is computed as

with q, the number of mapped classes. User’s accuracy of class i, noted U_i, is determined by

where p_ii represents the proportion of area mapped to class i that belongs to reference class i, and p_i., the total proportion of area mapped to class i. Producer’s accuracy is calculated as

where p_jj is the mapped proportion of area of class j that belongs to reference class j and p_.j is total proportion of area of reference class j.

Variance of overall accuracy (V (Ô)) (Eq. 5) is computed as

where q is the number of map classes, W_i is the proportion of area mapped to class i, Û_i is the class wise user’s accuracy and n_i is the number of sample units in stratum i.

Variance of user’s accuracy (V(Û_i)) is estimated according to Eq. 6:

and variance of producer’s accuracy of reference class j = k is estimated according to Eq. 7:

Where ({hat{N}}_{.j}) is equal to ({sum }_{i=1}^{q}frac{{N}_{i.}}{{n}_{i}.}{n}_{{ij}}) the estimate of marginal total number of pixels of reference class j; N_j., the marginal total of map class j, n_j. is the total number of sample units in class j and n_i. is the marginal total number of pixels in mapped class i. Standard errors for variance estimates can be obtained simply by taking the square root of the variances.

Finally, when a confusion matrix expressed in terms of area proportions is used, the proportion of area ({hat{p}}_{.k}) (Eq. 8) of mapped class k is expressed as:

The standard error (Sleft({hat{P}}_{.k}right)) (Eq. 9) of the estimated proportion of mapped class k is:

The associated area and standard error of area of class k, ({hat{A}}_{k}) and (Sleft({hat{A}}_{k}right)) are computed as Eq. 10 and Eq. 11:

where A is the total pixel counted area of the map. A 95% confidence interval can be obtained by multiplying (Sleft({hat{A}}_{k}right)) by 1.96.

We delivered final planted area estimates for winter wheat, other winter cereals, and rapeseed by mid-August 2022, mid-September 2023, and mid-June 2024 and mid-July 2025, respectively, to SSSU and FAO partners. The winter wheat planted area was derived as 85.7% of the total winter cereal areas. This proportion was computed from the report by the Ministry of Agricultural Policy and Food of Ukraine on winter crop planted areas for the 2022 season, released in December 2021 (SSSU, 2024). Standard error for winter wheat was adjusted in the same manner. From 2025 onward, we started using the previous season winter wheat to other winter cereals ratio, published in final SSSU statistics (State Statistics Service of Ukraine, 2024).

Remote Sensing imagery interpretation

The assignation of a sample unit, in our case pixels, or a cluster to a class is driven by a deductive approach with varying complexity based on the number of strata and their definitions. In a five classes setup, such as during the 2023 growing season, the labeler must first address the question Is this pixel cropland? The recognition of non-cropland classes is usually straightforward and will mostly be grassland, water, forested and built-up areas.

Once it is established that a sample unit belongs to cropland, the next question will be Does this sample unit belong to winter crops? Winter crops are typically planted in the autumn of the previous year and should exhibit a pre-winter green-up signal, appearing red in false color composites (as shown in the top row of Figs. 6, 7). Following an often cloudy and/or snowy winter period (illustrated in the second row of Figs. 6, 7), winter crops emerge from dormancy and resume growth. As demonstrated in the fourth row of Figs. 6, 7, a winter crop should be clearly photosynthetically active in April and reach peak Normalized Difference Vegetation Index (NDVI) values in May in Ukraine. We recommend using false color NIR-R-G composites for early season annotation as the NIR channel is more sensitive to the vegetation’s chlorophyll. If the observed sample unit meets the criteria for winter crops, the annotator should then carefully inspect the imagery in true color (R-G-B) mode, looking for the characteristic rapeseed flowering signal from mid-April to the end of May. Onset and peak flowering are visible in 2024-04-29 and 2024-05-13 composites in Fig. 6.

If no flowering is observed in this period, the annotator can safely conclude the sample unit is a winter cereal. Additionally, very high photosynthetic activity in the pre-winter period, is a non-conclusive hint that a winter crop might be rapeseed. Usually sown first, rapeseed has more time to establish than other winter crops, holds more chlorophyll and has therefore often higher NDVI values in autumn (see top row in Figs. 6, 7).

The last rows of Figs. 6, 7, cover the senescence to harvest phase of winter crops, usually starting in June. During this period, the annotator should expect a decrease in NDVI values as the crop dries, ultimately resulting in a clear harvest signal around end of July (visible in 2024-07-22 composite in Fig. 7, South from the sample unit).

Spring cereals can often be mistaken for winter cereals, particularly by less experienced annotators who may overlook the early season signals. Spring cereals exhibit a growth pattern similar to that of winter crops, peaking in May-June and maturing for harvest in July. However, unlike winter cereals, spring cereals are not planted in autumn and thus experience a delayed green-up, often revealing bare soils until early April, followed by a rapid green-up, as illustrated in Fig. 8.

If a crop is neither classified as non-crop, winter, nor spring crop, the labeler should then consider the question, Is this sample unit a summer crop or is it non-cultivated? In Ukraine, most summer crops display a bare soil signal until the end of May (rows 1 to 4 in Fig. 9), followed by rapid green-up in June and July (bottom row in Fig. 9). The annotator can further categorize non-cultivated cropland into two types: barren plots, which are often managed through tillage, typically do not exhibit significant vegetation signals throughout the growing season, as illustrated in Fig. 10.

Secondly, we define fallow as cropland covered in naturally grown vegetation. This vegetation greens up alongside winter crops but senesces in autumn with summer crops. When annotating grass-covered fallows, particularly those found along the frontline due to cropland abandonment, the annotator should look for “always green” plots, as shown in Fig. 11. Confusions can arise between fallow and cropped plots, especially during early June annotation. To mitigate this risk, we recommend “context-aware annotation,” as grassy fallows are often spatially co-located along the frontline, disrupting the typical winter crop/summer crop mosaic that characterizes Ukraine’s cropland. Fallow may be terminologically wrong, considering most weeded-in fields are, in reality, abandoned.

While labeling, the annotator may encounter sample units located at the edges between two classes. In such cases, we advise conducting a peer review to reach a consensual decision on whether the sample unit primarily represents one class or the other. Additionally, we recommend utilizing very high-resolution imagery, such as Google Satellite background maps, to aid in this decision-making process.