In-season cotton yield forecasting using high resolution satellite imagery

Authors:

Abstract
Accurate in-season cotton yield forecasting at fine spatial resolution remains a challenge due to the crop's physiological complexity, the limitations of UAS-based monitoring at scale, and the inadequacy of existing satellite-based methods that rely on single-date or magnitude-only vegetation indices. This study developed Pheno-SST, a hybrid deep learning architecture combining a Transformer encoder and a Mamba state-space branch. This model trained on near-daily PlanetScope NDVI time series and its first-order derivative on years 2021-2023 and tested on 2024 growth season in South Texas, USA. The first-order derivative, encoding the rate and direction of canopy change, was critical for cross year model generalization. Models trained on NDVI magnitude alone overfitted on the unseen 2024 test year, while Pheno-SST with derivative input achieved an RMSE of 2.89 and a near-zero bias of +0.30, outperforming all competing architectures including BiLSTM, PatchTST, Random Forest, TCN, and DLinear. Notably, the 2024 season was statistically distinct from training years in peak timing, canopy persistence, and cumulative growth, yet the model generalized effectively, demonstrating robustness to environmental heterogeneity. Conformal prediction intervals confirmed well-calibrated uncertainty, with 89.6% and 92.5% coverage at the 90% and 95% confidence levels, respectively. In-season forecasting showed that reliable predictions require the crop to reach approximately 105 to 117 days after planting, consistent with the onset of reproductive-stage dynamics in cotton. These results establish a scalable, physiologically informed, and uncertainty-aware framework for operational satellite-based cotton yield forecasting.

This publication has 42 references indexed in Scilit: