Deep time-series generative/predictive exploratory study of site-specific crop yield using satellite imagery

Abstract Purpose:: Pixel-level crop yield estimation using satellite imagery is a less explored problem with limited comparison of deep time-series models and their generalisability. Reliable prediction is constrained by sparse ground truth, spatial autocorrelation, imbalanced yield distributions, and temporal inconsistencies from cloud removal. We investigate how temporal alignment and feature pruning influence within-field interpolation versus spatial block generalisation. Furthermore, we introduce a new time-series generative augmentation pipeline to address yield imbalance and investigate its impact on prediction accuracy and robustness in a comparative predictive modelling framework. Methods:: Using 20,170 winter wheat pixel-year observations (2017--2021; 19 field-years) at Witzwil, Switzerland, we benchmark Random Forest (RF), a Convolutional Neural Network--Attention--Gated Recurrent Unit hybrid (CNN-A-GRU), and a selective state-space model (Mamba). We compare seven-pass Consistent Timestamps via Single Closest Pass (CTSCP) against 90-day seasonal averaging, prune features from $K=20$ to $1$, and introduce a regime-stratified temporal conditional generative adversarial network (T-CGAN) for extreme yield-bin augmentation. Results:: RF achieved the highest within-field accuracy ($R^2=0.901$), but its performance declined substantially on average across unseen spatial blocks within the same farm ($R^2=0.329$). In contrast, CNN-A-GRU demonstrated stronger average within-farm spatial generalisation ($R^2=0.444$), improving to $0.485$ with T-CGAN augmentation, while Mamba improved from $0.434$ to $0.464$. Consistent temporal sampling improved prediction accuracy across all models ($\Delta R^2\approx+0.022$), while feature pruning maintained competitive performance using only six variables. Conclusion:: High within-field accuracy does not guarantee spatial transferability; sequence models with discrete phenological sampling generalise better on average across unseen blocks within the same farm. Generative augmentation further improves average prediction accuracy under within-farm spatial distribution shift.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11569341/v1
Primary Topic
Remote Sensing in Agriculture
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Deep time-series generative/predictive exploratory study of site-specific crop yield using satellite imagery

A. Hornero, Eskandari Hamidreza, Sara Sharifzadeh, Xianghua Xie et al.
1 links
Research Square
Remote Sensing in Agriculture
preprint

Deep time-series generative/predictive exploratory study of site-specific crop yield using satellite imagery

A. Hornero, Eskandari Hamidreza, Sara Sharifzadeh, Xianghua Xie, Rosie Finnegan
preprint en

Abstract

Abstract Purpose:: Pixel-level crop yield estimation using satellite imagery is a less explored problem with limited comparison of deep time-series models and their generalisability. Reliable prediction is constrained by sparse ground truth, spatial autocorrelation, imbalanced yield distributions, and temporal inconsistencies from cloud removal. We investigate how temporal alignment and feature pruning influence within-field interpolation versus spatial block generalisation. Furthermore, we introduce a new time-series generative augmentation pipeline to address yield imbalance and investigate its impact on prediction accuracy and robustness in a comparative predictive modelling framework. Methods:: Using 20,170 winter wheat pixel-year observations (2017--2021; 19 field-years) at Witzwil, Switzerland, we benchmark Random Forest (RF), a Convolutional Neural Network--Attention--Gated Recurrent Unit hybrid (CNN-A-GRU), and a selective state-space model (Mamba). We compare seven-pass Consistent Timestamps via Single Closest Pass (CTSCP) against 90-day seasonal averaging, prune features from $K=20$ to $1$, and introduce a regime-stratified temporal conditional generative adversarial network (T-CGAN) for extreme yield-bin augmentation. Results:: RF achieved the highest within-field accuracy ($R^2=0.901$), but its performance declined substantially on average across unseen spatial blocks within the same farm ($R^2=0.329$). In contrast, CNN-A-GRU demonstrated stronger average within-farm spatial generalisation ($R^2=0.444$), improving to $0.485$ with T-CGAN augmentation, while Mamba improved from $0.434$ to $0.464$. Consistent temporal sampling improved prediction accuracy across all models ($\Delta R^2\approx+0.022$), while feature pruning maintained competitive performance using only six variables. Conclusion:: High within-field accuracy does not guarantee spatial transferability; sequence models with discrete phenological sampling generalise better on average across unseen blocks within the same farm. Generative augmentation further improves average prediction accuracy under within-farm spatial distribution shift.

Research Square
Swansea University (GB), Instituto de Agricultura Sostenible (ES)
Zero hunger
Remote Sensing in Agriculture
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon