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

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

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

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.

Dataset	Countries	Crops	Years	Fields	Pixel-Level	Resolution (Optical)		Features	Curated
						Spatial	Temporal
SwissYield [ 37 ]	1	2	2017–2021	73	✓	$10\text{\,}\mathrm{m}$	$\sim$ 5 days	14	✗
CropNet [ 22 ]	1	4	2017-2022	0	✗	$9\text{\,}\mathrm{k}\mathrm{m}$	$\sim$ 14 days	13	✗
YieldSAT (ours)	4	4	2016-2024	2,173	✓	$10\text{\,}\mathrm{m}$	$\sim$ 5 days	72	✓

Table 1: Comparison of the YieldSAT dataset with other crop yield prediction datasets.

YieldSAT	Fields					Pixels	Area (ha)		Years	S2 Images
	Corn	Rapeseed	Soybean	Wheat	Total		Total	Average
Argentina	185	✗	440	126	751	$\sim$ 5.3 M	$\sim$ 55,898	74.3	2017-2024	42,931
Brazil	118	✗	293	140	551	$\sim$ 4.2 M	$\sim$ 43,125	78.2	2017–2024	19,308
Uruguay	✗	✗	572	✗	572	$\sim$ 2.1 M	$\sim$ 32,804	57.3	2018–2022	24,318
Germany	✗	111	✗	188	299	$\sim$ 0.6 M	$\sim$ 6,460	21.6	2016–2022	26,998
Summary	303	111	1,305	454	2,173	$\sim$ 12.2 M	$\sim$ 138,288	57.8	2016-2024	113,555

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 ] )

Modality	Source	Product	Resolution
			Spatial	Temporal
Multispectral	Sentinel-2 L2A	B01 - Coastal Aerosol	$60\text{\,}\mathrm{m}$	$\thicksim$ 5 days
		B02 - Blue	$10\text{\,}\mathrm{m}$
		B03 - Green	$10\text{\,}\mathrm{m}$
		B04 - Red	$10\text{\,}\mathrm{m}$
		B05 - Red Edge 1	$20\text{\,}\mathrm{m}$

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).

Evaluation			Field-Level										Subfield (Pixel)-Level
Evaluation			ARG-S		BRA-C		GER-R		GER-W		URG-S		ARG-S		BRA-C		BRA-S		GER-R		GER-W		URG-S
Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
S2	✗	3D-ConvLSTM [ 13 ]	0.79	0.55	0.82	0.74	0.81	0.58	0.65	1.12	0.77	0.51	0.65	0.90	0.45	2.13	0.39	0.94	0.49	1.20	0.34	2.40	0.41	1.22
	✗	3D-LSTM [ 32 ]	0.77	0.58	0.82	0.74	0.82	0.57	0.54	1.28	0.73	0.56	0.65	0.90	0.46	2.13	0.39	0.93	0.48	1.20	0.30	2.46	0.39	1.23
	✗	LSTM [ 36 ]	0.72	0.64	0.75	0.88	0.62	0.83	0.55	2.60	0.66	0.62	0.60	0.96	0.42	2.20	0.34	0.98	0.36	1.33	0.32	2.47	0.37	1.26
	✗	Transformer [ 12 ]	0.73	0.63	0.79	0.82	0.75	0.67	0.56	1.26	0.72	0.56	0.62	0.94	0.44	2.16	0.38	0.94	0.44	1.25	0.32	2.43	0.38	1.24

Table 4: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation		Leave-One-Year-Out		Leave-One-Region-Out
Dataset	Model	RMSE ( $\downarrow$ ) t/ha	$R^{2}$ ( $\uparrow$ ) -	RMSE ( $\downarrow$ ) t/ha	$R^{2}$ ( $\uparrow$ ) -
Soybean	DE-LSTM	0.70 $\pm$ 0.25	0.55 $\pm$ 0.64	0.65 $\pm$ 0.20	0.65 $\pm$ 0.87
	DE-3D-LSTM	0.23 $\pm$ 0.11	0.63 $\pm$ 0.40	0.19 $\pm$ 0.14	0.73 $\pm$ 0.53
	Baseline LSTM	0.85 $\pm$ 0.23	0.50 $\pm$ 0.50	0.72 $\pm$ 0.15	0.64 $\pm$ 0.15
Corn	DE-LSTM	2.06 $\pm$ 0.82	0.46 $\pm$ 0.52	1.81 $\pm$ 0.86	0.59 $\pm$ 5.28
	DE-3D-LSTM	1.70 $\pm$ 0.61	0.63 $\pm$ 0.36	1.36 $\pm$ 0.83	0.76 $\pm$ 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.

Panel of Figure 9 — 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 Values	Wheat	Rapeseed	Soybean	Corn
Max. Yield (t/ha)	20	10	15	45
Standard Moisture (%)	15	9	15	16

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 $10\text{\,}\mathrm{m}$ grid for pixel-wise regression.

Figure 12: Label support size and standard deviation per $10\text{\,}\mathrm{m}$ 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.

Evaluation	Country	Crop
p-value

Appendix

Table 7: Kruskal–Wallis H-test between yield distributions grouped by countries and by crop type.

****=p<0.0001

Comparision	Argentina	Brazil	Germany	Uruguay
Argentina	-
Brazil		-	ns
Germany		ns	-
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\geq 0.05

*=p<0.05

**=p<0.01

***=p<0.001

****=p<0.0001

Comparision	Corn	Rapeseed	Soybean	Wheat
Corn	-
Rapeseed		-	ns
Soybean		ns	-
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\geq 0.05

*=p<0.05

**=p<0.01

***=p<0.001

****=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 .

Evaluation	Year	Region
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

Year	2017	2018	2019	2020	2021	2022
2017	-
2018		-	ns
2019		ns	-	**
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\geq 0.05

*=p<0.05

**=p<0.01

***=p<0.001

****=p<0.0001

Region	0	1	2	3	4	5	6	7
0	-						**
1		-
2			-
3				-
4					-	ns		ns
5					ns	-	ns	ns

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\geq 0.05

*=p<0.05

**=p<0.01

***=p<0.001

****=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.

Evaluation				Field-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
CV10	Sentinel-2	X	3D-ConvLSTM	0.84 $\pm$ 0.07	1.13 $\pm$ 0.22	0.79 $\pm$ 0.12	0.55 $\pm$ 0.16	0.92 $\pm$ 0.08	0.62 $\pm$ 0.24	0.82 $\pm$ 0.09	0.74 $\pm$ 0.12	0.76 $\pm$ 0.09	0.34 $\pm$ 0.05	0.73 $\pm$ 0.09	0.42 $\pm$ 0.16	0.81 $\pm$ 0.10	0.58 $\pm$ 0.19	0.65 $\pm$ 0.16	1.12 $\pm$ 0.34	0.77 $\pm$ 0.06	0.51 $\pm$ 0.10
		X	3D-LSTM	0.74 $\pm$ 0.15	1.44 $\pm$ 0.39	0.77 $\pm$ 0.11	0.58 $\pm$ 0.15	0.89 $\pm$ 0.10	0.70 $\pm$ 0.25	0.82 $\pm$ 0.11	0.74 $\pm$ 0.11	0.76 $\pm$ 0.08	0.34 $\pm$ 0.06	0.63 $\pm$ 0.16	0.48 $\pm$ 0.21	0.82 $\pm$ 0.09	0.57 $\pm$ 0.14	0.54 $\pm$ 0.26	1.28 $\pm$ 0.40	0.73 $\pm$ 0.09	0.56 $\pm$ 0.12
		X	LSTM	0.79 $\pm$ 0.11	1.29 $\pm$ 0.24	0.72 $\pm$ 0.12	0.64 $\pm$ 0.14	0.85 $\pm$ 0.10	0.84 $\pm$ 0.21	0.75 $\pm$ 0.2	0.88 $\pm$ 0.22	0.68 $\pm$ 0.12	0.39 $\pm$ 0.05	0.62 $\pm$ 0.16	0.49 $\pm$ 0.18	0.62 $\pm$ 0.25	0.83 $\pm$ 0.26	0.55 $\pm$ 0.41	2.60 $\pm$ 1.28	0.66 $\pm$ 0.11	0.62 $\pm$ 0.12
		X	Transformer	0.77 $\pm$ 0.07	1.35 $\pm$ 0.21	0.73 $\pm$ 0.11	0.63 $\pm$ 0.16	0.87 $\pm$ 0.11	0.78 $\pm$ 0.23	0.79 $\pm$ 0.19	0.82 $\pm$ 0.19	0.75 $\pm$ 0.12	0.35 $\pm$ 0.08	0.62 $\pm$ 0.12	0.49 $\pm$ 0.19	0.75 $\pm$ 0.14	0.67 $\pm$ 0.21	0.56 $\pm$ 0.33	1.26 $\pm$ 0.24	0.72 $\pm$ 0.10	0.56 $\pm$ 0.10

Appendix

Table 13: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation				Subfield (Pixel)-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE
CV10	Sentinel-2	X	3D-ConvLSTM	0.65 $\pm$ 0.11	2.2 $\pm$ 0.27	0.65 $\pm$ 0.06	0.9 $\pm$ 0.11	0.79 $\pm$ 0.1	1.1 $\pm$ 0.18	0.45 $\pm$ 0.12	2.13 $\pm$ 0.31	0.39 $\pm$ 0.07	0.94 $\pm$ 0.06	0.23 $\pm$ 0.05	1.38 $\pm$ 0.18	0.49 $\pm$ 0.1	1.2 $\pm$ 0.22	0.34 $\pm$ 0.09	2.4 $\pm$ 0.34	0.41 $\pm$ 0.06	1.22 $\pm$ 0.12
		X	3D-LSTM	0.59 $\pm$ 0.13	2.37 $\pm$ 0.38	0.65 $\pm$ 0.07	0.9 $\pm$ 0.11	0.79 $\pm$ 0.11	1.08 $\pm$ 0.2	0.46 $\pm$ 0.14	2.13 $\pm$ 0.37	0.39 $\pm$ 0.07	0.93 $\pm$ 0.06	0.24 $\pm$ 0.05	1.37 $\pm$ 0.19	0.48 $\pm$ 0.12	1.2 $\pm$ 0.23	0.3 $\pm$ 0.2	2.46 $\pm$ 0.43	0.39 $\pm$ 0.07	1.23 $\pm$ 0.12
		X	LSTM	0.59 $\pm$ 0.07	2.4 $\pm$ 0.26	0.6 $\pm$ 0.08	0.96 $\pm$ 0.1	0.74 $\pm$ 0.12	1.21 $\pm$ 0.19	0.42 $\pm$ 0.13	2.2 $\pm$ 0.35	0.34 $\pm$ 0.09	0.98 $\pm$ 0.06	0.22 $\pm$ 0.06	1.39 $\pm$ 0.17	0.36 $\pm$ 0.14	1.33 $\pm$ 0.23	-0.32 $\pm$ 0.64	3.38 $\pm$ 1.04	0.37 $\pm$ 0.07	1.26 $\pm$ 0.12
		X	Transformer	0.61 $\pm$ 0.05	2.33 $\pm$ 0.23	0.62 $\pm$ 0.08	0.94 $\pm$ 0.1	0.76 $\pm$ 0.12	1.17 $\pm$ 0.22	0.44 $\pm$ 0.13	2.16 $\pm$ 0.31	0.38 $\pm$ 0.08	0.94 $\pm$ 0.07	0.22 $\pm$ 0.06	1.39 $\pm$ 0.17	0.44 $\pm$ 0.1	1.25 $\pm$ 0.24	0.32 $\pm$ 0.14	2.43 $\pm$ 0.27	0.38 $\pm$ 0.08	1.24 $\pm$ 0.13

Appendix

Table 14: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation				Field-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
LOYO	Sentinel-2	X	3D-ConvLSTM	0.61 $\pm$ 0.55	1.75 $\pm$ 0.51	0.65 $\pm$ 0.31	0.71 $\pm$ 0.18	0.78 $\pm$ 0.88	1.00 $\pm$ 0.40	0.20 $\pm$ 0.32	1.58 $\pm$ 0.75	0.36 $\pm$ 0.42	0.55 $\pm$ 0.15	-0.06 $\pm$ 1.30	0.82 $\pm$ 0.42	0.43 $\pm$ 0.24	1.02 $\pm$ 0.23	0.23 $\pm$ 0.60	1.67 $\pm$ 0.60	0.63 $\pm$ 0.25	0.65 $\pm$ 0.15
		X	3D-LSTM	0.53 $\pm$ 0.49	1.93 $\pm$ 0.71	0.66 $\pm$ 0.24	0.70 $\pm$ 0.21	0.77 $\pm$ 1.94	1.02 $\pm$ 0.37	0.36 $\pm$ 0.66	1.42 $\pm$ 0.42	0.12 $\pm$ 0.86	0.65 $\pm$ 0.22	-0.33 $\pm$ 1.03	0.92 $\pm$ 0.56	0.41 $\pm$ 0.23	1.03 $\pm$ 0.34	0.00 $\pm$ 0.74	1.90 $\pm$ 0.72	0.62 $\pm$ 0.30	0.66 $\pm$ 0.17
		X	LSTM	0.46 $\pm$ 0.59	2.06 $\pm$ 0.54	0.50 $\pm$ 0.22	0.85 $\pm$ 0.25	0.77 $\pm$ 1.70	1.02 $\pm$ 0.30	0.19 $\pm$ 1.50	1.59 $\pm$ 0.48	0.06 $\pm$ 1.66	0.67 $\pm$ 0.21	-0.05 $\pm$ 1.61	0.82 $\pm$ 0.40	-0.07 $\pm$ 0.77	1.39 $\pm$ 0.46	-1.74 $\pm$ 2.65	3.14 $\pm$ 1.32	0.61 $\pm$ 0.23	0.67 $\pm$ 0.12
		X	Transformer	0.57 $\pm$ 0.64	1.83 $\pm$ 0.33	0.52 $\pm$ 0.21	0.84 $\pm$ 0.20	0.78 $\pm$ 0.94	0.99 $\pm$ 0.51	0.19 $\pm$ 0.61	1.59 $\pm$ 0.59	0.28 $\pm$ 0.64	0.59 $\pm$ 0.16	-0.09 $\pm$ 1.04	0.83 $\pm$ 0.42	0.33 $\pm$ 0.44	1.10 $\pm$ 0.30	0.14 $\pm$ 0.91	1.76 $\pm$ 0.50	0.60 $\pm$ 0.33	0.68 $\pm$ 0.16

Appendix

Table 15: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation				Subfield (Pixel)-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
LOYO	Sentinel-2	X	3D-ConvLSTM	0.48 $\pm$ 0.29	2.68 $\pm$ 0.33	0.54 $\pm$ 0.18	1.04 $\pm$ 0.16	0.67 $\pm$ 0.58	1.38 $\pm$ 0.32	0.15 $\pm$ 0.11	2.67 $\pm$ 0.90	0.19 $\pm$ 0.11	1.08 $\pm$ 0.17	0.06 $\pm$ 0.19	1.52 $\pm$ 0.45	0.21 $\pm$ 0.15	1.49 $\pm$ 0.31	0.15 $\pm$ 0.27	2.71 $\pm$ 0.79	0.34 $\pm$ 0.09	1.28 $\pm$ 0.12
		X	3D-LSTM	0.47 $\pm$ 0.25	2.70 $\pm$ 0.39	0.55 $\pm$ 0.14	1.03 $\pm$ 0.18	0.68 $\pm$ 0.5	1.35 $\pm$ 0.32	0.21 $\pm$ 0.19	2.56 $\pm$ 0.73	0.13 $\pm$ 0.21	1.12 $\pm$ 0.18	0.03 $\pm$ 0.24	1.54 $\pm$ 0.53	0.18 $\pm$ 0.26	1.52 $\pm$ 0.43	0.01 $\pm$ 0.3	2.93 $\pm$ 0.91	0.35 $\pm$ 0.11	1.27 $\pm$ 0.14
		X	LSTM	0.39 $\pm$ 0.23	2.90 $\pm$ 0.27	0.46 $\pm$ 0.16	1.12 $\pm$ 0.17	0.61 $\pm$ 0.73	1.49 $\pm$ 0.46	0.15 $\pm$ 0.25	2.66 $\pm$ 0.66	0.04 $\pm$ 0.44	1.18 $\pm$ 0.22	0.08 $\pm$ 0.17	1.50 $\pm$ 0.46	-0.05 $\pm$ 0.35	1.71 $\pm$ 0.28	-0.68 $\pm$ 0.77	3.82 $\pm$ 1.21	0.33 $\pm$ 0.08	1.29 $\pm$ 0.1
		X	Transformer	0.42 $\pm$ 0.35	2.83 $\pm$ 0.29	0.48 $\pm$ 0.14	1.10 $\pm$ 0.16	0.67 $\pm$ 0.62	1.38 $\pm$ 0.41	0.19 $\pm$ 0.16	2.60 $\pm$ 0.8	0.16 $\pm$ 0.20	1.10 $\pm$ 0.14	0.05 $\pm$ 0.19	1.53 $\pm$ 0.45	0.14 $\pm$ 0.19	1.55 $\pm$ 0.32	0.13 $\pm$ 0.30	2.74 $\pm$ 0.70	0.33 $\pm$ 0.11	1.29 $\pm$ 0.12

Appendix

Table 16: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation				Field-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
LORO	Sentinel-2	X	3D-ConvLSTM	0.68 $\pm$ 1.89	1.59 $\pm$ 0.48	0.69 $\pm$ 0.12	0.67 $\pm$ 0.12	0.80 $\pm$ 0.21	0.95 $\pm$ 0.24	0.55 $\pm$ 0.74	1.19 $\pm$ 0.36	0.37 $\pm$ 1.63	0.55 $\pm$ 0.18	0.47 $\pm$ 0.30	0.58 $\pm$ 0.24	0.25 $\pm$ 0.53	1.16 $\pm$ 0.31	0.14 $\pm$ 1.05	1.76 $\pm$ 0.75	0.68 $\pm$ 0.52	0.61 $\pm$ 0.30
		X	3D-LSTM	0.49 $\pm$ 2.60	2.00 $\pm$ 0.73	0.70 $\pm$ 0.13	0.67 $\pm$ 0.14	0.81 $\pm$ 0.21	0.92 $\pm$ 0.16	0.59 $\pm$ 1.15	1.13 $\pm$ 0.49	0.41 $\pm$ 1.89	0.53 $\pm$ 0.15	0.48 $\pm$ 0.27	0.57 $\pm$ 0.23	0.26 $\pm$ 0.30	1.16 $\pm$ 0.36	-0.08 $\pm$ 1.35	1.98 $\pm$ 1.09	0.65 $\pm$ 0.90	0.63 $\pm$ 0.26
		X	LSTM	0.47 $\pm$ 1.36	2.05 $\pm$ 0.75	0.64 $\pm$ 0.15	0.72 $\pm$ 0.15	0.7 $\pm$ 0.38	1.18 $\pm$ 0.26	0.39 $\pm$ 1.80	1.39 $\pm$ 0.37	0.28 $\pm$ 2.07	0.59 $\pm$ 0.21	0.27 $\pm$ 0.19	0.68 $\pm$ 0.23	-0.16 $\pm$ 0.65	1.45 $\pm$ 0.49	-2.71 $\pm$ 3.03	3.66 $\pm$ 1.26	0.63 $\pm$ 0.43	0.65 $\pm$ 0.31
		X	Transformer	0.55 $\pm$ 2.12	1.88 $\pm$ 0.46	0.59 $\pm$ 0.17	0.77 $\pm$ 0.14	0.71 $\pm$ 0.46	1.16 $\pm$ 0.33	0.35 $\pm$ 0.29	1.43 $\pm$ 0.59	0.33 $\pm$ 1.85	0.57 $\pm$ 0.17	0.37 $\pm$ 0.28	0.63 $\pm$ 0.25	0.17 $\pm$ 0.84	1.22 $\pm$ 0.39	0.01 $\pm$ 1.89	1.89 $\pm$ 0.69	0.63 $\pm$ 1.11	0.65 $\pm$ 0.21

Appendix

Table 17: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Evaluation				Subfield (Pixel)-Level
Evaluation				ARG-C		ARG-S		ARG-W		BRA-C		BRA-S		BRA-W		GER-R		GER-W		URG-S
Splitting	Modalities	Fusion Method	Model	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE	$R^{2}$	RMSE
LORO	Sentinel-2	X	3D-ConvLSTM	0.54 $\pm$ 0.38	2.52 $\pm$ 0.33	0.57 $\pm$ 0.11	1.01 $\pm$ 0.10	0.69 $\pm$ 0.24	1.33 $\pm$ 0.23	0.33 $\pm$ 0.14	2.35 $\pm$ 0.37	0.19 $\pm$ 0.22	1.08 $\pm$ 0.33	0.16 $\pm$ 0.07	1.44 $\pm$ 0.31	0.12 $\pm$ 0.29	1.57 $\pm$ 0.30	0.07 $\pm$ 0.43	2.84 $\pm$ 0.69	0.36 $\pm$ 0.19	1.26 $\pm$ 0.29
		X	3D-LSTM	0.39 $\pm$ 0.50	2.91 $\pm$ 0.56	0.59 $\pm$ 0.10	0.98 $\pm$ 0.10	0.71 $\pm$ 0.24	1.30 $\pm$ 0.18	0.34 $\pm$ 0.20	2.33 $\pm$ 0.53	0.22 $\pm$ 0.16	1.06 $\pm$ 0.34	0.20 $\pm$ 0.07	1.41 $\pm$ 0.30	0.17 $\pm$ 0.20	1.52 $\pm$ 0.45	0.02 $\pm$ 0.43	2.92 $\pm$ 0.94	0.36 $\pm$ 0.21	1.26 $\pm$ 0.28
		X	LSTM	0.44 $\pm$ 0.28	2.79 $\pm$ 0.35	0.55 $\pm$ 0.10	1.03 $\pm$ 0.09	0.61 $\pm$ 0.32	1.48 $\pm$ 0.24	0.27 $\pm$ 0.06	2.46 $\pm$ 0.44	0.16 $\pm$ 0.21	1.10 $\pm$ 0.35	0.11 $\pm$ 0.09	1.48 $\pm$ 0.34	-0.05 $\pm$ 0.33	1.71 $\pm$ 0.54	-0.78 $\pm$ 1.18	3.93 $\pm$ 1.16	0.34 $\pm$ 0.17	1.28 $\pm$ 0.30
		X	Transformer	0.43 $\pm$ 0.40	2.80 $\pm$ 0.40	0.52 $\pm$ 0.10	1.05 $\pm$ 0.09	0.61 $\pm$ 0.37	1.49 $\pm$ 0.30	0.27 $\pm$ 0.12	2.47 $\pm$ 0.50	0.21 $\pm$ 0.18	1.07 $\pm$ 0.32	0.11 $\pm$ 0.10	1.48 $\pm$ 0.32	0.09 $\pm$ 0.49	1.59 $\pm$ 0.31	0.10 $\pm$ 0.68	2.79 $\pm$ 0.61	0.31 $\pm$ 0.27	1.31 $\pm$ 0.26

Appendix

Table 18: Results for the RMSE (t/ha) (

\downarrow

) and the

R^{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.

Crop Yield Prediction

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

See all 18 papers

Apr 21, 2026

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

Crop yield prediction is one of the most important challenge, which is crucial to world food security and policy-making decisions. The conventional forecasting techniques are limited in their accuracy with reference to the fact that they utilize static data sources that do not reflect the dynamic and intricate relationships that exist between the variables of the environment over time [5,13]. This paper presents Attention-Based Multi-Modal Deep Learning Framework (ABMMDLF), which is suggested to be used in high-accuracy spatio-temporal crop yield prediction. The model we use combines multi-year satellite imagery, high-resolution time-series of meteorological data and initial soil properties as opposed to the traditional models which use only one of the aforementioned factors [12, 21]. The main architecture involves the use of Convolutional Neural Networks (CNN) to extract spatial features and a Temporal Attention Mechanism to adaptively weight important phenological periods targeted by the algorithm to change over time and condition on spatial features of images and video sequences. As can be experimentally seen, the proposed research work provides an R^2 score of 0.89, which is far better than the baseline models do.

Gopal Krishna Shyam, Ila Chandrakar

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

Agricultural Remote Sensing Spatiotemporal Forecasting Crop Yield Prediction

Ask PDF

Sep 16, 2026

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

Reliable agricultural yield statistics are typically reported at coarse administrative scales, whereas modern geospatial machine learning methods require spatially explicit, pixel level supervision. This mismatch has limited the development of large-scale benchmarks for crop yield learning using multimodal Earth observation data. A reproducible benchmark, AgroBench, is presented for transforming publicly available U.S. county level crop yield statistics into weakly supervised pixel-level crop time series. Each crop pixel time series is paired with a county-level yield value as a weak supervisory signal rather than a directly measured pixel-level yield label. Our geospatial data generation pipeline integrates USDA crop yield statistics with crop-specific land cover masks, Sentinel 2 multispectral imagery, Sentinel-1 synthetic aperture radar observations, climatic variables, and terrain information to produce temporally aligned multimodal sequences describing individual crop pixels throughout the growing season. The resulting benchmark contains over 13 million observations from 788,654 unique crop pixels spanning 5,107 county year combinations across eight growing seasons (2017 to 2024) for five major U.S. crops. To facilitate standardized evaluation, we establish a crop yield prediction benchmark using a Leave-One-Year-Out evaluation protocol and provide baseline results using representative machine learning models. By releasing the complete data generation pipeline, benchmark dataset, and evaluation protocol, AgroBench provides a reproducible foundation for future research in weakly supervised learning, multimodal remote sensing, spatiotemporal modeling, and geospatial foundation models for agriculture.

Udaiveer Singh, Rajiv Ranjan, Shashank Tamaskar +1

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

Multi-Angular Remote Sensing Remote Sensing Benchmarks Crop Yield Prediction Agricultural Remote Sensing Weakly Supervised Learning Remote Sensing

Ask PDF

Mar 7, 2026

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

This paper proposes a new method for crop yield prediction, which is essential for developing management strategies, informing insurance assessments, and ensuring long-term food security. Although existing data-driven approaches have shown promise in this domain, their performance often degrades when applied across large geographic regions and long time periods. This limitation arises from two key challenges: (1) difficulty in jointly capturing short-term and long-term temporal patterns, and (2) inability to effectively accommodate spatial data variability in agricultural systems. Ignoring these issues often leads to unreliable predictions for specific regions or years, which ultimately affects policy decisions and resource allocation. In this paper, we propose a new predictive framework to address these challenges. First, we introduce a new backbone model architecture that captures both short-term daily-scale crop growth dynamics and long-term dependencies across years. To further improve generalization across diverse spatial regions, we augment this model with a retrieval-based adaptation strategy. Recognizing the substantial yield variation across years, we design a novel retrieval-and-refinement pipeline that adjusts retrieved samples by removing cross-year bias not explained by input features. Our experiments on real-world county-level corn yield data over 630 counties in the US demonstrate that our method consistently outperforms different types of baselines. The results also verify the effectiveness of the retrieval-based augmentation method in improving model robustness under spatial heterogeneity.

Yiming Sun, Qi Cheng, Licheng Liu +3

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

Crop Yield Prediction Spatiotemporal Forecasting Retrieval-Augmented Time Series Forecasting Multiscale Time-Series Modeling Precision Agriculture

Ask PDF

Abstract

Crop Yield Prediction

Explore similar work

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

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

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