cs.CVApr 1, 2026

YieldSAT: A Multimodal Benchmark Dataset for High-Resolution Crop Yield Prediction

Authors: Miro Miranda, Deepak Pathak, Patrick Helber, Benjamin Bischke, Hiba Najjar, Francisco Mena, Cristhian Sanchez, Akshay Pai, 

Organizations: RPTU Kaiserslautern-Landau · DFKI GmbH · Vision Impulse GmbH · University of Groningen

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Crop yield prediction requires substantial data to train scalable models. However, creating yield prediction datasets is constrained by high acquisition costs, heterogeneous data quality, and data privacy regulations. Consequently, existing datasets are scarce, low in quality, or limited to regional levels or single crop types, hindering the development of scalable data-driven solutions. In this work, we release YieldSAT, a large, high-quality, and multimodal dataset for high-resolution crop yield prediction. YieldSAT spans various climate zones across multiple countries, including Argentina, Brazil, Uruguay, and Germany, and includes major crop types, including corn, rapeseed, soybeans, and wheat, across 2,173 expert-curated fields. In total, over 12.2 million yield samples are available, each with a spatial resolution of 10 m. Each field is paired with multispectral satellite imagery, resulting in 113,555 labeled satellite images, complemented by auxiliary environmental data. We demonstrate the potential of large-scale and high-resolution crop yield prediction as a pixel regression task by comparing various deep learning models and data fusion architectures. Furthermore, we highlight open challenges arising from severe distribution shifts in the ground truth data under real-world conditions. To mitigate this, we explore a domain-informed Deep Ensemble approach that exhibits significant performance gains. The dataset is available at https://yieldsat.github.io/.

Figures & tables

Figure 1: Schematic overview of the YieldSAT dataset and the data collection and preprocessing. At harvest, a combine harvester collects point data containing various information (yield, geolocation, time, and moisture), referred to as a yield map. For each yield map, EO data is collected (satellite imagery, weather, soil, and topography data). The target yield map and the input data are preprocessed into an ML-ready data format, formulating yield prediction as a pixel-wise regression task.
DatasetCountriesCropsYearsFieldsPixel-LevelResolution (Optical)FeaturesCurated
      SpatialTemporal  
SwissYield [ 37 ]122017–202173✓10 m10\text{\,}\mathrm{m}∼\sim 5 days14✗
CropNet [ 22 ]142017-20220✗9 km9\text{\,}\mathrm{k}\mathrm{m}∼\sim 14 days13✗
YieldSAT (ours)442016-20242,173✓10 m10\text{\,}\mathrm{m}∼\sim 5 days72✓
Table 1: Comparison of the YieldSAT dataset with other crop yield prediction datasets.
YieldSATFieldsPixelsArea (ha)YearsS2 Images
 CornRapeseedSoybeanWheatTotal TotalAverage  
Argentina185✗440126751∼\sim 5.3 M∼\sim 55,89874.32017-202442,931
Brazil118✗293140551∼\sim 4.2 M∼\sim 43,12578.22017–202419,308
Uruguay✗✗572✗572∼\sim 2.1 M∼\sim 32,80457.32018–202224,318
Germany✗111✗188299∼\sim 0.6 M∼\sim 6,46021.62016–202226,998
Summary3031111,3054542,173∼\sim 12.2 M∼\sim 138,28857.82016-2024113,555
Table 2: Overview of the available ground truth yield data. In total, 2,173 labeled fields (yield images) are available.
Figure 2: Spatial distribution of the collected yield data for each region and crop type. The yield data is colored by crop type. Left: South America, right: Europe. The marker size indicates the field size. (Map source: [ 26 ] )
ModalitySourceProductResolution
   SpatialTemporal
MultispectralSentinel-2 L2AB01 - Coastal Aerosol60 m60\text{\,}\mathrm{m}∼\thicksim 5 days
B02 - Blue10 m10\text{\,}\mathrm{m}
B03 - Green10 m10\text{\,}\mathrm{m}
B04 - Red10 m10\text{\,}\mathrm{m}
B05 - Red Edge 120 m20\text{\,}\mathrm{m}
Table 3: Overview of all available data modalities in the YieldSat dataset and their characteristics.
Figure 3: Example satellite time series of a soybean field from seeding (left) to harvesting (right). The last image shows the collected yield at harvest.
Figure 4: Yield data distribution plots for each crop type and country, averaged per field.
Figure 5: Examples of rasterized yield maps with the curation for each quality level ( good , average , and bad ).
Figure 6: t-SNE plot of the S2 surface reflectance colored by countries (left) and crops (right).
EvaluationField-LevelSubfield (Pixel)-Level
ARG-SBRA-CGER-RGER-WURG-SARG-SBRA-CBRA-SGER-RGER-WURG-S
ModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
S2✗3D-ConvLSTM [ 13 ]0.790.550.820.740.810.580.651.120.770.510.650.900.452.130.390.940.491.200.342.400.411.22
✗3D-LSTM [ 32 ]0.770.580.820.740.820.570.541.280.730.560.650.900.462.130.390.930.481.200.302.460.391.23
✗LSTM [ 36 ]0.720.640.750.880.620.830.552.600.660.620.600.960.422.200.340.980.361.330.322.470.371.26
✗Transformer [ 12 ]0.730.630.790.820.750.670.561.260.720.560.620.940.442.160.380.940.441.250.322.430.381.24
Table 4: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets. The best model is highlighted in bold, and the best overall score is underlined. ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
Figure 7: Qualitative results for a single soybean field from Argentina, generated with the 3D-LSTM model. Top left to bottom right: ground truth yield map, predicted yield map, prediction over target plot, input distribution over time, clipped relative pixel-wise error (100%), relative pixel-wise error (full range), histogram of predicted (blue) and target (green) values.
EvaluationLeave-One-Year-OutLeave-One-Region-Out
DatasetModelRMSE ( ↓\downarrow ) t/haR2R^{2} ( ↑\uparrow ) -RMSE ( ↓\downarrow ) t/haR2R^{2} ( ↑\uparrow ) -
SoybeanDE-LSTM0.70 ±\pm 0.250.55 ±\pm 0.640.65 ±\pm 0.200.65 ±\pm 0.87
DE-3D-LSTM0.23 ±\pm 0.110.63 ±\pm 0.400.19 ±\pm 0.140.73 ±\pm 0.53
Baseline LSTM0.85 ±\pm 0.230.50 ±\pm 0.500.72 ±\pm 0.150.64 ±\pm 0.15
CornDE-LSTM2.06 ±\pm 0.820.46 ±\pm 0.521.81 ±\pm 0.860.59 ±\pm 5.28
DE-3D-LSTM1.70 ±\pm 0.610.63 ±\pm 0.361.36 ±\pm 0.830.76 ±\pm 2.31
Table 5: Overview for crop yield prediction at the field level using temporal (Leave-One-Year-Out) and spatial splitting (Leave-One-Region-Out). All models, except the baseline, are defined as a Deep Ensemble [ 20 ] with 5 ensemble members. All models were trained on S2 data only on subsets from Argentina. DE = Deep Ensemble. The best score is highlighted in bold.
Figure 8: t-SNE of the network parameters for the Deep Ensemble model for different CV scenarios. Left: standard 10-CV, center: LOYO, right: LORO. The weights are colored by fold.
Figure 9: Visualization of the weight space diversity during model training. Left: Cosine similarity between two ensemble members during training. Right: PCA plot of the weight space during training, together with the loss. The trajectory in the weight space is highlighted in red from start to end.
Figure 10: t-SNE of the network parameters for the Deep Ensemble model trained with prior knowledge (3D-LSTM) under distribution shift. Left: LOYO, right: LORO.
Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Standard ValuesWheatRapeseedSoybeanCorn
Max. Yield (t/ha)20101545
Standard Moisture (%)1591516
Appendix
Table 6: Maximum accepted yield values in t/ha for every crop type and the standard moisture.
Figure 11: Schematic overview of data collection and preprocessing: A combine harvester collects point vector data of the yield, which is preprocessed and rasterized to align with S2 10 m10\text{\,}\mathrm{m} grid for pixel-wise regression.
Figure 12: Label support size and standard deviation per 10 m10\text{\,}\mathrm{m} pixel caused by the rasterization. Top row: sample count, bottom row: standard deviation.
Figure 13: Boxplots for pixel point count (left) and mean standard deviation (right) per crop and country.
Figure 14: Overall distribution of yield map quality over the entire dataset.
Figure 15: Overall distribution of yield map quality over the entire dataset grouped by country.
Figure 16: Overall distribution of yield map quality over the entire dataset grouped by country.
EvaluationCountryCrop
p-value
Appendix
Table 7: Kruskal–Wallis H-test between yield distributions grouped by countries and by crop type. ∗∗∗∗=p<0.0001****=p<0.0001 .
ComparisionArgentinaBrazilGermanyUruguay
Argentina-
Brazil-ns
Germanyns-
Uruguay-
Appendix
Table 8: Pairwise post-hoc comparisons of the yield distributions between individual countries . Each cell displays the statistical significance level of the difference between two countries based on the Dunn’s test using the Holm–Bonferroni correction. ns = no significance ( p≥0.05p\geq 0.05 ), ∗=p<0.05*=p<0.05 , ∗∗=p<0.01**=p<0.01 , ∗∗∗=p<0.001***=p<0.001 , ∗∗∗∗=p<0.0001****=p<0.0001 .
ComparisionCornRapeseedSoybeanWheat
Corn-
Rapeseed-ns
Soybeanns-
Wheat-
Appendix
Table 9: Pairwise post-hoc comparisons of the yield distributions between individual crops . Each cell displays the statistical significance level of the difference between two crops based on the Dunn’s test using the Holm–Bonferroni correction. ns = no significance ( p≥0.05p\geq 0.05 ), ∗=p<0.05*=p<0.05 , ∗∗=p<0.01**=p<0.01 , ∗∗∗=p<0.001***=p<0.001 , ∗∗∗∗=p<0.0001****=p<0.0001 .
Figure 17: Visualization of the surface reflectance and yield data distribution only for soybean in Argentina, grouped by years and regions. Top left: t-SNE plot of the S2 surface reflectance data grouped by years . Top right: t-SNE plot of the S2 surface reflectance data grouped by region . Bottom left: Kernel density estimation plot for the target yield data grouped by years . Bottom right: Kernel density estimation plot for the target yield data grouped by regions .
EvaluationYearRegion
p-value
Appendix
Table 10: Kruskal–Wallis H-test between yield distributions for soybean in Argentina, grouped by years and by regions. ∗∗∗∗=p<0.0001****=p<0.0001 .
Year201720182019202020212022
2017-
2018-ns
2019ns-**
2020**-
2021-
2022-
Appendix
Table 11: Pairwise post-hoc comparisons of the yield distributions for soybean in Argentina between individual years . Each cell displays the statistical significance level of the difference between two years based on Dunn’s test using the Holm–Bonferroni correction. ns = no significance ( p≥0.05p\geq 0.05 ), ∗=p<0.05*=p<0.05 , ∗∗=p<0.01**=p<0.01 , ∗∗∗=p<0.001***=p<0.001 , ∗∗∗∗=p<0.0001****=p<0.0001 .
Region01234567
0-**
1-
2-
3-
4-nsns
5ns-nsns
Appendix
Table 12: Pairwise post-hoc comparisons of the yield distributions for soybean in Argentina between individual regions . Each cell displays the statistical significance level of the difference between two regions based on Dunn’s test using the Holm–Bonferroni correction. ns = no significance ( p≥0.05p\geq 0.05 ), ∗=p<0.05*=p<0.05 , ∗∗=p<0.01**=p<0.01 , ∗∗∗=p<0.001***=p<0.001 , ∗∗∗∗=p<0.0001****=p<0.0001 .
Figure 18: Visualization of the weight space diversity during model training. Top: Cosine similarity between pairs of ensemble members during training. Bottom: PCA plot of the weight space during training, together with the loss. The trajectory in the weight space is highlighted in red from start to end.
Figure 19: Example ground truth and predicted yield maps together with the pixel-wise error for the standard CV, the LORO, and LOYO CV under distribution shift. The results are generated with the LSTM model [ 36 ] , trained on S2 data.
EvaluationField-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
CV10Sentinel-2X3D-ConvLSTM0.84 ±\pm 0.071.13 ±\pm 0.220.79 ±\pm 0.120.55 ±\pm 0.160.92 ±\pm 0.080.62 ±\pm 0.240.82 ±\pm 0.090.74 ±\pm 0.120.76 ±\pm 0.090.34 ±\pm 0.050.73 ±\pm 0.090.42 ±\pm 0.160.81 ±\pm 0.100.58 ±\pm 0.190.65 ±\pm 0.161.12 ±\pm 0.340.77 ±\pm 0.060.51 ±\pm 0.10
X3D-LSTM0.74 ±\pm 0.151.44 ±\pm 0.390.77 ±\pm 0.110.58 ±\pm 0.150.89 ±\pm 0.100.70 ±\pm 0.250.82 ±\pm 0.110.74 ±\pm 0.110.76 ±\pm 0.080.34 ±\pm 0.060.63 ±\pm 0.160.48 ±\pm 0.210.82 ±\pm 0.090.57 ±\pm 0.140.54 ±\pm 0.261.28 ±\pm 0.400.73 ±\pm 0.090.56 ±\pm 0.12
XLSTM0.79 ±\pm 0.111.29 ±\pm 0.240.72 ±\pm 0.120.64 ±\pm 0.140.85 ±\pm 0.100.84 ±\pm 0.210.75 ±\pm 0.20.88 ±\pm 0.220.68 ±\pm 0.120.39 ±\pm 0.050.62 ±\pm 0.160.49 ±\pm 0.180.62 ±\pm 0.250.83 ±\pm 0.260.55 ±\pm 0.412.60 ±\pm 1.280.66 ±\pm 0.110.62 ±\pm 0.12
XTransformer0.77 ±\pm 0.071.35 ±\pm 0.210.73 ±\pm 0.110.63 ±\pm 0.160.87 ±\pm 0.110.78 ±\pm 0.230.79 ±\pm 0.190.82 ±\pm 0.190.75 ±\pm 0.120.35 ±\pm 0.080.62 ±\pm 0.120.49 ±\pm 0.190.75 ±\pm 0.140.67 ±\pm 0.210.56 ±\pm 0.331.26 ±\pm 0.240.72 ±\pm 0.100.56 ±\pm 0.10
Appendix
Table 13: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the standard 10-fold cross-validation at the field level. ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
EvaluationSubfield (Pixel)-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2RMSER2RMSER2RMSER2RMSER2RMSER2RMSER2RMSER2RMSER2RMSE
CV10Sentinel-2X3D-ConvLSTM0.65 ±\pm 0.112.2 ±\pm 0.270.65 ±\pm 0.060.9 ±\pm 0.110.79 ±\pm 0.11.1 ±\pm 0.180.45 ±\pm 0.122.13 ±\pm 0.310.39 ±\pm 0.070.94 ±\pm 0.060.23 ±\pm 0.051.38 ±\pm 0.180.49 ±\pm 0.11.2 ±\pm 0.220.34 ±\pm 0.092.4 ±\pm 0.340.41 ±\pm 0.061.22 ±\pm 0.12
X3D-LSTM0.59 ±\pm 0.132.37 ±\pm 0.380.65 ±\pm 0.070.9 ±\pm 0.110.79 ±\pm 0.111.08 ±\pm 0.20.46 ±\pm 0.142.13 ±\pm 0.370.39 ±\pm 0.070.93 ±\pm 0.060.24 ±\pm 0.051.37 ±\pm 0.190.48 ±\pm 0.121.2 ±\pm 0.230.3 ±\pm 0.22.46 ±\pm 0.430.39 ±\pm 0.071.23 ±\pm 0.12
XLSTM0.59 ±\pm 0.072.4 ±\pm 0.260.6 ±\pm 0.080.96 ±\pm 0.10.74 ±\pm 0.121.21 ±\pm 0.190.42 ±\pm 0.132.2 ±\pm 0.350.34 ±\pm 0.090.98 ±\pm 0.060.22 ±\pm 0.061.39 ±\pm 0.170.36 ±\pm 0.141.33 ±\pm 0.23-0.32 ±\pm 0.643.38 ±\pm 1.040.37 ±\pm 0.071.26 ±\pm 0.12
XTransformer0.61 ±\pm 0.052.33 ±\pm 0.230.62 ±\pm 0.080.94 ±\pm 0.10.76 ±\pm 0.121.17 ±\pm 0.220.44 ±\pm 0.132.16 ±\pm 0.310.38 ±\pm 0.080.94 ±\pm 0.070.22 ±\pm 0.061.39 ±\pm 0.170.44 ±\pm 0.11.25 ±\pm 0.240.32 ±\pm 0.142.43 ±\pm 0.270.38 ±\pm 0.081.24 ±\pm 0.13
Appendix
Table 14: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the standard 10-fold cross-validation at the subfield (pixel)-level. ARG = Argentina, URG = Uruguay, GER = Germany, BRA = Brazil. S = soybean, R = rapeseed, W = wheat, C = Corn.
EvaluationField-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
LOYOSentinel-2X3D-ConvLSTM0.61 ±\pm 0.551.75 ±\pm 0.510.65 ±\pm 0.310.71 ±\pm 0.180.78 ±\pm 0.881.00 ±\pm 0.400.20 ±\pm 0.321.58 ±\pm 0.750.36 ±\pm 0.420.55 ±\pm 0.15-0.06 ±\pm 1.300.82 ±\pm 0.420.43 ±\pm 0.241.02 ±\pm 0.230.23 ±\pm 0.601.67 ±\pm 0.600.63 ±\pm 0.250.65 ±\pm 0.15
X3D-LSTM0.53 ±\pm 0.491.93 ±\pm 0.710.66 ±\pm 0.240.70 ±\pm 0.210.77 ±\pm 1.941.02 ±\pm 0.370.36 ±\pm 0.661.42 ±\pm 0.420.12 ±\pm 0.860.65 ±\pm 0.22-0.33 ±\pm 1.030.92 ±\pm 0.560.41 ±\pm 0.231.03 ±\pm 0.340.00 ±\pm 0.741.90 ±\pm 0.720.62 ±\pm 0.300.66 ±\pm 0.17
XLSTM0.46 ±\pm 0.592.06 ±\pm 0.540.50 ±\pm 0.220.85 ±\pm 0.250.77 ±\pm 1.701.02 ±\pm 0.300.19 ±\pm 1.501.59 ±\pm 0.480.06 ±\pm 1.660.67 ±\pm 0.21-0.05 ±\pm 1.610.82 ±\pm 0.40-0.07 ±\pm 0.771.39 ±\pm 0.46-1.74 ±\pm 2.653.14 ±\pm 1.320.61 ±\pm 0.230.67 ±\pm 0.12
XTransformer0.57 ±\pm 0.641.83 ±\pm 0.330.52 ±\pm 0.210.84 ±\pm 0.200.78 ±\pm 0.940.99 ±\pm 0.510.19 ±\pm 0.611.59 ±\pm 0.590.28 ±\pm 0.640.59 ±\pm 0.16-0.09 ±\pm 1.040.83 ±\pm 0.420.33 ±\pm 0.441.10 ±\pm 0.300.14 ±\pm 0.911.76 ±\pm 0.500.60 ±\pm 0.330.68 ±\pm 0.16
Appendix
Table 15: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the Leave-One-Year-Out cross-validation at the field level. ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
EvaluationSubfield (Pixel)-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
LOYOSentinel-2X3D-ConvLSTM0.48 ±\pm 0.292.68 ±\pm 0.330.54 ±\pm 0.181.04 ±\pm 0.160.67 ±\pm 0.581.38 ±\pm 0.320.15 ±\pm 0.112.67 ±\pm 0.900.19 ±\pm 0.111.08 ±\pm 0.170.06 ±\pm 0.191.52 ±\pm 0.450.21 ±\pm 0.151.49 ±\pm 0.310.15 ±\pm 0.272.71 ±\pm 0.790.34 ±\pm 0.091.28 ±\pm 0.12
X3D-LSTM0.47 ±\pm 0.252.70 ±\pm 0.390.55 ±\pm 0.141.03 ±\pm 0.180.68 ±\pm 0.51.35 ±\pm 0.320.21 ±\pm 0.192.56 ±\pm 0.730.13 ±\pm 0.211.12 ±\pm 0.180.03 ±\pm 0.241.54 ±\pm 0.530.18 ±\pm 0.261.52 ±\pm 0.430.01 ±\pm 0.32.93 ±\pm 0.910.35 ±\pm 0.111.27 ±\pm 0.14
XLSTM0.39 ±\pm 0.232.90 ±\pm 0.270.46 ±\pm 0.161.12 ±\pm 0.170.61 ±\pm 0.731.49 ±\pm 0.460.15 ±\pm 0.252.66 ±\pm 0.660.04 ±\pm 0.441.18 ±\pm 0.220.08 ±\pm 0.171.50 ±\pm 0.46-0.05 ±\pm 0.351.71 ±\pm 0.28-0.68 ±\pm 0.773.82 ±\pm 1.210.33 ±\pm 0.081.29 ±\pm 0.1
Transformer0.42 ±\pm 0.352.83 ±\pm 0.290.48 ±\pm 0.141.10 ±\pm 0.160.67 ±\pm 0.621.38 ±\pm 0.410.19 ±\pm 0.162.60 ±\pm 0.80.16 ±\pm 0.201.10 ±\pm 0.140.05 ±\pm 0.191.53 ±\pm 0.450.14 ±\pm 0.191.55 ±\pm 0.320.13 ±\pm 0.302.74 ±\pm 0.700.33 ±\pm 0.111.29 ±\pm 0.12
Appendix
Table 16: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the Leave-One-Region-Out (LORO) cross-validation at the field level.ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
EvaluationField-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
LOROSentinel-2X3D-ConvLSTM0.68 ±\pm 1.891.59 ±\pm 0.480.69 ±\pm 0.120.67 ±\pm 0.120.80 ±\pm 0.210.95 ±\pm 0.240.55 ±\pm 0.741.19 ±\pm 0.360.37 ±\pm 1.630.55 ±\pm 0.180.47 ±\pm 0.300.58 ±\pm 0.240.25 ±\pm 0.531.16 ±\pm 0.310.14 ±\pm 1.051.76 ±\pm 0.750.68 ±\pm 0.520.61 ±\pm 0.30
X3D-LSTM0.49 ±\pm 2.602.00 ±\pm 0.730.70 ±\pm 0.130.67 ±\pm 0.140.81 ±\pm 0.210.92 ±\pm 0.160.59 ±\pm 1.151.13 ±\pm 0.490.41 ±\pm 1.890.53 ±\pm 0.150.48 ±\pm 0.270.57 ±\pm 0.230.26 ±\pm 0.301.16 ±\pm 0.36-0.08 ±\pm 1.351.98 ±\pm 1.090.65 ±\pm 0.900.63 ±\pm 0.26
XLSTM0.47 ±\pm 1.362.05 ±\pm 0.750.64 ±\pm 0.150.72 ±\pm 0.150.7 ±\pm 0.381.18 ±\pm 0.260.39 ±\pm 1.801.39 ±\pm 0.370.28 ±\pm 2.070.59 ±\pm 0.210.27 ±\pm 0.190.68 ±\pm 0.23-0.16 ±\pm 0.651.45 ±\pm 0.49-2.71 ±\pm 3.033.66 ±\pm 1.260.63 ±\pm 0.430.65 ±\pm 0.31
XTransformer0.55 ±\pm 2.121.88 ±\pm 0.460.59 ±\pm 0.170.77 ±\pm 0.140.71 ±\pm 0.461.16 ±\pm 0.330.35 ±\pm 0.291.43 ±\pm 0.590.33 ±\pm 1.850.57 ±\pm 0.170.37 ±\pm 0.280.63 ±\pm 0.250.17 ±\pm 0.841.22 ±\pm 0.390.01 ±\pm 1.891.89 ±\pm 0.690.63 ±\pm 1.110.65 ±\pm 0.21
Appendix
Table 17: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the Leave-One-Region-Out (LORO) cross-validation at the field level. ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
EvaluationSubfield (Pixel)-Level
ARG-CARG-SARG-WBRA-CBRA-SBRA-WGER-RGER-WURG-S
SplittingModalitiesFusion MethodModelR2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSER2R^{2}RMSE
LOROSentinel-2X3D-ConvLSTM0.54 ±\pm 0.382.52 ±\pm 0.330.57 ±\pm 0.111.01 ±\pm 0.100.69 ±\pm 0.241.33 ±\pm 0.230.33 ±\pm 0.142.35 ±\pm 0.370.19 ±\pm 0.221.08 ±\pm 0.330.16 ±\pm 0.071.44 ±\pm 0.310.12 ±\pm 0.291.57 ±\pm 0.300.07 ±\pm 0.432.84 ±\pm 0.690.36 ±\pm 0.191.26 ±\pm 0.29
X3D-LSTM0.39 ±\pm 0.502.91 ±\pm 0.560.59 ±\pm 0.100.98 ±\pm 0.100.71 ±\pm 0.241.30 ±\pm 0.180.34 ±\pm 0.202.33 ±\pm 0.530.22 ±\pm 0.161.06 ±\pm 0.340.20 ±\pm 0.071.41 ±\pm 0.300.17 ±\pm 0.201.52 ±\pm 0.450.02 ±\pm 0.432.92 ±\pm 0.940.36 ±\pm 0.211.26 ±\pm 0.28
XLSTM0.44 ±\pm 0.282.79 ±\pm 0.350.55 ±\pm 0.101.03 ±\pm 0.090.61 ±\pm 0.321.48 ±\pm 0.240.27 ±\pm 0.062.46 ±\pm 0.440.16 ±\pm 0.211.10 ±\pm 0.350.11 ±\pm 0.091.48 ±\pm 0.34-0.05 ±\pm 0.331.71 ±\pm 0.54-0.78 ±\pm 1.183.93 ±\pm 1.160.34 ±\pm 0.171.28 ±\pm 0.30
XTransformer0.43 ±\pm 0.402.80 ±\pm 0.400.52 ±\pm 0.101.05 ±\pm 0.090.61 ±\pm 0.371.49 ±\pm 0.300.27 ±\pm 0.122.47 ±\pm 0.500.21 ±\pm 0.181.07 ±\pm 0.320.11 ±\pm 0.101.48 ±\pm 0.320.09 ±\pm 0.491.59 ±\pm 0.310.10 ±\pm 0.682.79 ±\pm 0.610.31 ±\pm 0.271.31 ±\pm 0.26
Appendix
Table 18: Results for the RMSE (t/ha) ( ↓\downarrow ) and the R2R^{2} -score ( ↑\uparrow ) for different models and datasets for the Leave-One-Region-Out (LORO) cross-validation at the subfield (pixel)-level. ARG = Argentina, BRA = Brazil, GER = Germany, URG = Uruguay. C = corn, R = rapeseed, S = soybean, W = wheat.
More in this topic

Crop Yield Prediction

Crop yield prediction estimates harvests from crop, weather, soil, satellite, and management data to support agricultural planning.

See all 18 papers

Explore similar work

CardsList

  1. Attention-based Multi-modal Deep Learning Model of Spatio-temporal Crop Yield Prediction with Satellite, Soil and Climate Data

    Apr 21, 2026Gopal Krishna Shyam, Ila ChandrakarAgricultural Remote SensingSpatiotemporal Forecasting

    Presidency University, Bengaluru, India · University of Europe for Applied Sciences, Berlin

  2. AgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations

    Sep 16, 2026Udaiveer Singh, Rajiv Ranjan, Shashank Tamaskar +1Multi-Angular Remote SensingRemote Sensing Benchmarks

    Plaksha University, Mohali, Punjab, India · Purdue University, West Lafayette, Indiana, USA

  3. Retrieval-Augmented Multi-scale Framework for County-Level Crop Yield Prediction Across Large Regions

    Mar 7, 2026Yiming Sun, Qi Cheng, Licheng Liu +3Crop Yield PredictionSpatiotemporal Forecasting

    University of Pittsburgh, Pittsburgh, PA, USA · University of Wisconsin-Madison, Madison, WI, USA · University of Alabama, Tuscaloosa, AL, USA +1