Estimating village-level potato yields using satellite remote sensing and ground data in Uttar Pradesh: An ensemble approach

Introduction

Potato is a crucial staple crop in Uttar Pradesh, India, contributing significantly to the state's agricultural economy and food security. As the largest potato-producing state in India, accounting for over 35% of the national potato yield, accurate and timely prediction of potato production is essential for effective market management, crop planning, and ensuring food availability for the rapidly growing population. With its high carbohydrate content potato is considered a leading food ingredient and is one of the most popular and widely used vegetable in the world. It is the fourth largest food crop in the world following the Maize, wheat and rice respectively. Potato is consumed by more than one billion people all over the world and is grown in more than 100 countries in the world with a production of around 364 million tons in the year 2012. China is now the biggest potato producer and almost one-third of all potatoes is harvested in China and India. At present, potato is high quality vegetable cum food crop in respect of world food scenario. Existing methods for potato yield estimation in Uttar Pradesh primarily rely on manual field surveys, which are labourintensive, costly, and often suffer from delays in data collection and processing. Consequently, there is a need for a more efficient and reliable approach to potato yield prediction that can provide granular, village-level estimates. Recent advancements in remote sensing technology and artificial intelligence offer promising solutions to address this challenge. Remote sensing data from satellites, such as Sentinel-1 and Sentinel-2, provide detailed, multi-temporal information on crop conditions, which, when combined with machine learning algorithms, can enable accurate and scalable yield forecasting. (Lin et al., 2023) (Mahdi et al., 2020). It's important to note that while ground yield are valuable, they are subject to certain limitations, such as sampling errors and potential biases if the selected plots are not truly representative. Modern techniques like remote sensing and crop modeling are increasingly being integrated with ground yield to improve the accuracy and efficiency of yield estimation. The use of remote sensing and modeling for yield estimation. While these sources don't specifically address Ground Yield methods in Uttar Pradesh, they highlight the broader context of yield estimation and the role of technology in improving its accuracy.

Objectives

 To estimate village-level potato yields by using multisource remote sensing data with meteorological and groundtruth data.  To evaluate the impact of agro-meteorological factors and biophysical parameters on potato yield variability, and to analyze their interactions with crop phenology. [1,2,4,7] . Frequent data collection can help mitigate issues caused by cloud cover and other temporal gaps, improving the robustness of yield predictions (Awad, 2019) [5] . The normalized difference vegetation index (NDVI) and the soil adjusted vegetation index (SAVI) are commonly used and have shown good correlation with actual yields (Al-Gaadi et al., 2016; Bala & Islam, 2009) [1,4] . Different vegetation indices may perform better depending on the specific conditions and crop stages. For example, NDVI, LAI, and fPAR have shown high correlation coefficients (R² values) in various studies (Bala & Islam, 2009; Huang & Han, 2014) [4,6] . Data collected during specific growth stages, such as the tuber initiation stage, can be more predictive of final yields. Earlyseason data often provide better correlations with marketable yields (Li et al., 2021) [3] . The timing of data collection (e.g., July, August, September) can significantly impact the accuracy of yield predictions, with some models performing better with data from specific months (Gómez et al., 2019) [2] . Different machine learning models, such as Random Forest Regression (RFR), Support Vector Regression (SVR), and Regression Quantile Lasso, have varying levels of accuracy. Feature selection and preprocessing steps are crucial for improving model performance (Gómez et al., 2019; Li et al., 2021) [2,3] . Models that incorporate feature selection to reduce multicollinearity among predictors tend to perform better. For instance, the Regression Quantile Lasso and Leap Backwards models showed lower RMSE and higher R² values when correlated predictors were removed (Gómez et al., 2019) [2] .

Incorporating cultivar-specific data can significantly improve the accuracy of yield predictions. Different cultivars may respond differently to environmental conditions, and including this information helps tailor the models more precisely (Li et al., 2021) [3] . Soil fertility and climate conditions (e.g., precipitation, temperature) also affect the accuracy of yield predictions.

Studies have shown that correlations between yield and remote sensing data vary with soil types and climate systems (Huang & Han, 2014) [6] . [2] Random Forest (RF) Uses Sentinel 2 bands and Potato Productivity Index (PPI); better than NDVI. RMSE: 15.42%, R²: 0.77 (Gómez et al., 2021) [9] Multiple Linear Regression (MLR) Uses agronomic, phytophenological, and meteorological data. MAPE: <15% (Piekutowska et al., 2021) [11] The table above compares various machine learning models used for potato yield estimation. Models like Regression Quantile Lasso and Leap Backwards show high accuracy with feature selection, while Support Vector Machine Radial performs well without it. Random Forest models using the Potato Productivity Index outperform those using NDVI. ANFIS model demonstrates strong predictive capabilities, with ANFIS showing superior performance due to its fuzzy logic integration.

Materials and Methods

Study Area

Uttar Pradesh, India's largest potato-producing state, plays a crucial role in the country's agricultural economy. Among its prominent agricultural districts, Kannauj stands out for its significant contribution to potato farming. Located in the fertile Indo-Gangetic Plain, Kannauj spans approximately 1,993 square kilometres and lies between 26°54'N to 27°06'N latitude and 79°44'E to 80°01'E longitude. The district comprises three tehsils and over 716 villages. It is bordered by Farrukhabad to the west, Kanpur Nagar to the south, and Hardoi to the east, forming a vital agricultural hub. The district is known for cultivating key crops such as potato, wheat, rice, mustard, and sugarcane, with potato being particularly noteworthy due to its contribution to Uttar Pradesh's dominance in national potato production. The fertile alluvial soils and reliable irrigation systems, including canals and tube wells, provide ideal conditions for intensive agriculture. Kannauj experiences a subtropical climate, with temperatures ranging from 5 °C in winter to 45 °C in summer. The potatogrowing season is primarily during the winter months from October to February. The district receives an average annual rainfall of 800-1,000 mm, mostly during the monsoon season, with supplemental irrigation supporting rabi crops like potato. Kannauj was chosen as the study area for this research because of its high potato productivity.

Ground Yield Data

Yield data was collected at the time of crop harvesting for validation purposes. The traditional Crop Cutting Experiment (CCE) methodology was not applied in this approach. Instead, a Smart Sampling method was used to select fields based on satellite remote sensing-based proxy yield parameters and crop health indicators such as NDVI, LSWI, and FAPAR, supported by ground truth data.

In the Kannauj district, selected fields were physically measured, and the entire field area was considered for yield assessment. During harvesting, the crop was harvested and threshed, and both biomass and grain weight were recorded from the actual measured area to calculate yield for model validation.

Crop Mask

Accurate identification of the crop sown area is essential for reliable yield estimation, as it ensures that the analysis is confined to relevant agricultural zones. In this study, the potato crop mask was provided by the Department of Technology, Star Agribazaar Technology Pvt. Ltd. Spanning an area of 66,091 hectares, the mask was developed using time-series Sentinel-2 satellite imagery at a spatial resolution of 10 meters for the Rabi season of 2024-25. This high-resolution crop mask precisely delineates regions under potato cultivation and serves as a foundational dataset for data preprocessing and yield estimation. The crop mask overlaying village boundary is illustrated in Figure 5, showcasing the spatial distribution of potato fields across the study area. Leaf Area Index: LAI (Leaf Area Index) derived from Sentinel-2 imagery using vegetation indices and radiative transfer models provides critical insights into crop canopy structure and photosynthetic activity. Sentinel-2's high-resolution red and near-infrared (NIR) bands enable accurate LAI estimation.

Frequent revisit times ensure timely monitoring of potato growth, biomass production, and stress conditions, supporting effective crop management and productivity enhancement.

LAI=a⋅VI+b

Where: VI is the vegetation index, and a and b are coefficients determined through calibration with field data. Soil Moisture (Sentinel-1)-Soil moisture can be estimated using Sentinel-1 Synthetic Aperture Radar (SAR) data due to its sensitivity to surface moisture content. Sentinel-1's dualpolarized (VV and VH) C-band radar provides valuable information on soil conditions, even under cloudy or low-light conditions. Regular soil moisture mapping with Sentinel-1 supports monitoring of potato fields, aiding in irrigation scheduling, drought assessment, and optimizing water use efficiency. Its high temporal resolution ensures frequent updates, enabling precise management of agricultural practices to enhance productivity and sustainability.

SM=α⋅ln(σVV0)+β

Where: α\alphaα and β\betaβ are coefficients obtained through calibration with in-situ soil moisture data.

Meteorological Data

Daily rainfall and temperature data for the Rabi 2024-25 season were sourced from the Indian Meteorological Department (IMD). These datasets included minimum and maximum temperatures, which are critical for assessing thermal stress and its impact on crop growth and development. Rainfall data provided insights into precipitation patterns, essential for evaluating water availability during the crop's phenological stages. This meteorological information was integrated with satellite-derived parameters to enhance the accuracy of potato yield estimation by accounting for environmental variability.

Methods

The following methodology chart (Figure 6) outlines the step-by-step approach employed in this study for estimating village level potato yields. These features serve as the input for model development, where a multi-source ensemble approach is used to build predictive models for potato yield estimation. The models are then validated by comparing the predicted results with actual ground data. Finally, the Village level yield estimation is carried out, using the validated models to estimate the potato yield at a fine spatial resolution. This comprehensive processing workflow allows for accurate yield predictions and effective agricultural management.

Results and Discussion

Interactions potato yield with satellite derived biophysical parameters during different crop phenological stages

The study evaluated the temporal dynamics of satellite-derived biophysical parameters such as NDVI, LSWI, FAPAR, and LAI for yield categories-good, moderate, and poor-across different phenological stages. Line charts were used to depict the variation of these indices over time. These temporal profiles provide insights into how different growth conditions influence yield. However, the results also revealed that some indices, particularly at peak growth stages, reached saturation or displayed non-linear correlations with yield, deviating from theoretical expectations. This highlights the need to account for index saturation and integrate additional parameters for more accurate yield modeling. Good-yield locations consistently exhibited higher NDVI, LSWI, FAPAR, and LAI values, indicating robust canopy growth, water availability, and photosynthetic activity. Moderate-yield locations showed lower peaks, while poor-yield areas demonstrated subdued profiles, reflecting stress factors like limited water or suboptimal growing conditions. However, it was observed that these indices sometimes reached saturation points or did not directly correlate with yield as expected, indicating potential limitations in their standalone predictive capacity. The estimated yields, however, have a mean of 308.58 Q/ha, which is very close to the ground truth mean, suggesting that the model has accurately predicted the overall yield. The standard deviation of the estimated yields is 14.18 Q/ha, which is slightly higher than the grounds yields, indicating that the model has captured the core distribution of the yields with slightly reduced variability. This suggests that the model has high accuracy in predicting the yields, with a narrower spread in the predicted values compared to the actual data. The slight reduction in variability for the predicted yields reflects the model's consistency and its ability to predict yield trends with high precision, closely matching the observed field data. This indicates that the model is highly effective and accurate in estimating potato yields, with minimal error and a high degree of reliability in predicting yield variations. The validation of predicted yield against ground-collected yield data demonstrated an excellent correlation, with a coefficient of determination (R 2 ) of 0.85. This high R 2 value indicates a strong agreement between the predicted and observed yields, validating the accuracy and reliability of the prediction model for yield estimation.

Validation

Conclusion

Estimating potato yields accurately requires a comprehensive approach that incorporates cultivar-specific information, as yields are significantly influenced by tuber size and cultivar type. Solely relying on satellite-derived data often falls short, emphasizing the need for integrating additional agronomic and ground-level data. Previous studies also underscore the importance of these supplementary inputs for improving prediction accuracy. This study highlights that potato yield estimation using remote sensing is influenced by various factors, including the resolution and type of satellite data, the selection of vegetation indices, the timing of data acquisition, and the machine learning models used. Incorporating critical factors like cultivar differences and soil conditions further enhances model reliability. By optimizing these components, yield prediction accuracy can be significantly improved, offering valuable insights for agricultural planning, market regulation, and food security. This integrative approach lays a foundation for scalable, data-driven solutions in agricultural management.

References

1. Al-Gaadi K, Hassaballa A, Tola E, Kayad A, Madugundu R, Alblewi B, et al. Prediction of potato crop yield using precision agriculture techniques. PLoS ONE. 2016;11. https://doi.org/10.1371/journal.pone.0162219
2. Gómez D, Salvador P, Sanz J, Casanova J. Potato yield prediction using machine learning techniques and Sentinel-2 data. Remote Sens. 2019;11:1745. https://doi.org/10.3390/RS11151745
3. Li D, Miao Y, Gupta S, Rosen C, Yuan F, Wang C, et al. Improving potato yield prediction by combining cultivar information and UAV remote sensing data using machine learning. Remote Sens. 2021;13:3322. https://doi.org/10.3390/rs13163322
4. Bala S, Islam A. Correlation between potato yield and MODIS-derived vegetation indices. Int J Remote Sens. 2009;30:2491-2507. https://doi.org/10.1080/01431160802552744
5. Awad M. Toward precision in crop yield estimation using remote sensing and optimization techniques. Agriculture. 2019. https://doi.org/10.3390/AGRICULTURE9030054
6. Huang J, Han D. Meta-analysis of influential factors on crop yield estimation by remote sensing. Int J Remote Sens. 2014;35:2267-2295. https://doi.org/10.1080/01431161.2014.890761
7. Vallentin C, Harfenmeister K, Itzerott S, Kleinschmit B, Conrad C, Spengler D. Suitability of satellite remote sensing data for yield estimation in northeast Germany. Precision Agric. 2021;23:52-82. https://doi.org/10.1007/s11119-021-09827-6
8. Yu T, Zhou J, Fan J, Wang Y, Zhang Z. Potato leaf area index estimation using multi-sensor unmanned aerial vehicle (UAV) imagery and machine learning. Remote Sens. 2023. https://doi.org/10.3390/rs15164108
9. Gómez D, Salvador P, Sanz J, Casanova J. New spectral indicator Potato Productivity Index based on Sentinel-2 data to improve potato yield prediction: a machine learning approach. Int J Remote Sens. 2021;42:3426-3444. https://doi.org/10.1080/01431161.2020.1871102
10. Mhango J, Harris W, Monaghan J. Relationships between the spatio-temporal variation in reflectance data from the Sentinel-2 satellite and potato (Solanum tuberosum L.) yield and stem density. Remote Sens. 2021;13:4371. https://doi.org/10.3390/rs13214371
11. Piekutowska M, Niedbała G, Piskier T, Lenartowicz T, Pilarski K, Wojciechowski T, et al. The application of multiple linear regression and artificial neural network models for yield prediction of very early potato cultivars before harvest. Agronomy. 2021;11:885. https://doi.org/10.3390/AGRONOMY11050885
12. Khoshnevisan B, Rafiee S, Omid M, Mousazadeh H. Prediction of potato yield based on energy inputs using multi-layer adaptive neuro-fuzzy inference system. Measurement. 2014;47:521-530. https://doi.org/10.1016/J.MEASUREMENT.2013.09.020
13. Fortin J, Anctil F, Parent L, Bolinder M. Site-specific early season potato yield forecast by neural network in Eastern Canada. Precision Agric. 2011;12:905-923. https://doi.org/10.1007/s11119-011-9233-6
14. Salvador P, Gómez D, Sanz J, Casanova J. Estimation of potato yield using satellite data at a municipal level: a machine learning approach. ISPRS Int J Geo-Inf. 2020;9:343. https://doi.org/10.3390/ijgi9060343
15. Gupta A, Sarkar K, Bhattacharya D, Dhakre D. Potato yield modeling based on meteorological factors using discriminant analysis and artificial neural networks. Int J Vegetable Sci. 2021;28:465-476. https://doi.org/10.1080/19315260.2021.2021342