Context-aware deep learning model for yield prediction in potato using time-series UAS multispectral data

Authors: YADAV, SURAJ - Mississippi State University; ZHANG, XIN - Mississippi State University; WIJEWARDANE, NUWAN - Mississippi State University; Feldman, Maximilian; QIN, RUIJUN - Oregon State University; Huang, Yanbo; SAMIAPPAN, SATHISHKUMAR - Mississippi State University; Young, Wyatt; Gonzalez Tapia, Francisco

Submitted to: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/3/2025
Publication Date: 2/5/2025
Citation: Yadav, S.A., Zhang, X., Wijewardane, N.K., Feldman, M.J., Qin, R., Huang, Y., Samiappan, S., Young, W., Gonzalez Tapia, F. 2025. Context-aware deep learning model for yield prediction in potato using time-series UAS multispectral data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 18:6096-6115. https://doi.org/10.1109/JSTARS.2025.3539217.
DOI: https://doi.org/10.1109/JSTARS.2025.3539217

Interpretive Summary: Potato yield is difficult to predict given that the economic product, the tuber, is hidden beneath the soil surface; and that crop productivity is highly dependent upon cultivar, management, and climate. Use of drone-based, multispectral imaging provides direct measurements of plant size and reflectance characteristics at high resolution across time and space which has been demonstrated to improve yield prediction ability in other crops. Scientists at the USDA-ARS Temperate Tree Fruit and Vegetable Research Unit, the Forage Seed and Cereal Research Unit in Prosser, WA, and the Genetics and Sustainable Agriculture Research Unit in Mississippi State, MS in collaboration with researchers at Mississippi State University and Oregon State University performed a study to evaluate the capability of drone imaging to predict potato yield. Four different russet potato cultivars were evaluated under diverse nitrogen fertilization rates (0 to 639 kg/ha) during a two-year period. Drone imaging was performed on a weekly basis throughout the entire field season and was used in a regression model to predict total yield from destructive harvest at the end of each field season. Results from this study suggest that features extracted from drone images confer substantial capacity to predict tuber yield, particularly when imaging data can be collected throughout the entire field season.

Technical Abstract: The study demonstrated the efficacy of integrating time-series uncrewed aerial system (UAS) multispectral imaging with data-driven deep learning methodologies to systematically and precisely predict field-scale crop yield throughout the growing seasons. A UAS equipped with a MicaSense RedEdge MX+ sensor was used for data acquisition at the Hermiston Agricultural Research and Extension Center, Oregon State University. The data were collected throughout the potato (Solanum tuberosum L.) growing seasons under varied nitrogen (N)-rates ranging from 0 to 639 kg/ha. The raw data was preprocessed using Pix4Dmapper and the quantum geographic information system (QGIS). A linear unmixing model followed by Otsu-based adaptive auto-segmentation was implemented to generate soil-masked spatio-spectral fusion maps for accurate vegetation feature extraction. The proposed feature engineering and prediction model followed a two-fold approach: (1) adoption of partial least squares regression (PLSR) algorithm to extract features relevant to yield, and (2) a novel Context-Aware attention and Residual connection Convolution-Bidirectional gated recurrent unit Bidirectional Long Short-Term Memory-Network (CAR Conv1D-BiGRU-BiLSTM-Net) to exploit time-series multifeatures information to predict final yield. On integrating the PLSR-derived robust features, the proposed model demonstrated an increase in predictive capability from emergence (T1) to bulking (T4) growth stage by effectively capturing the temporal dynamics of physiological and biological traits. Overall, using multi-features such as simple ratio (SR), Chlorophyll Green (CHLGR), Modified Anthocyanin Reflectance Index (MARI), vegetation fraction (Vf), and N-rate from T1-T4 growth stage resulted in predictive accuracy with high R2 = 0.775 and low RMSE of 16.4%, outperforming other deep learning models. The source code is available at https://github.com/SAY70/CAR-CNN-BiRNN.