High accuracy inside the field, collapse across it: random forest drops from R² 0.90 to 0.329 on unseen blocks, while sequence models hold near 0.44
Synopsis
Pixel-level winter wheat yield prediction at one Swiss farm shows that high within-field interpolation accuracy does not carry over to unseen blocks. Random forest reached R² about 0.90 under a coordinate-grouped split but averaged 0.329 under contiguous within-year spatial block cross-validation, while CNN-A-GRU and Mamba held 0.444 and 0.434, rising to 0.485 and 0.464 with regime-stratified generative augmentation. Seven-pass phenological sampling also beat 90-day seasonal averaging for all three models.
Interpretation
Moving to blocks of the same farm that were not used in training sharply lowers accuracy, and how far it falls depends on how a model organizes a season of observations. Under a coordinate-grouped split, adjacent 10 m pixels share weather, sowing dates and soil, so random forest reached R² about 0.90 for within-field interpolation; such splits were previously treated as evidence of model capability. On 20,170 winter wheat pixel-year observations from 19 field-years between 2017 and 2021, 5-fold contiguous within-year spatial block cross-validation gave random forest a training R² of 0.984 and a mean test R² of 0.329, with −0.313 in Fold 5, where the train–test yield divergence reached a Kolmogorov–Smirnov statistic of 0.2438.
Models that treat the season as an ordered sequence are more stable on average across unseen blocks of the same farm. Random forest led on within-field interpolation but ranked last on the spatial block test, while the sequence models fit training blocks less tightly yet tested better. CNN-A-GRU averaged a test R² of 0.444 and Mamba 0.434, both above the 0.329 of random forest; by fold, CNN-A-GRU reached 0.611 in Fold 2 and 0.623 in Fold 4.
Selecting single satellite passes aligned to phenology preserves more usable signal than compressing a year into three 90-day means. Fixed calendar windows and coarse seasonal composites average out brief stages such as stem elongation and flowering, a common practice in optical remote sensing yield modelling. Against 90-day seasonal averaging, seven consistent timestamps raised R² by about 0.022 for random forest, CNN-A-GRU and Mamba, with Welch test p-values far below 0.001 for all three.
Synthesizing time series for extreme yield bins lowers error at both tails without changing overall accuracy. Extreme low- and high-yield samples number only tens to a few hundred real training pixels per fold, and time-series augmentation targeted at these bins previously lacked a controlled comparison. Under the within-field split, random forest RMSE in the 0–2 t/ha bin fell from 0.409 to 0.382 t/ha at 100% augmentation volume, and in the 12–14 t/ha bin from 0.866 to 0.839 t/ha at 50% volume; global R² changes were not significant for any model, p≥0.258.
Perspective
The findings rest on 20,170 ten-metre pixel-year observations from 19 winter wheat field-years at Witzwil, Switzerland, between 2017 and 2021, and assess transfer across contiguous blocks within one farm rather than across farms or agro-ecosystems. The within-field benchmark uses 8,581 pixels with complete seven-pass sequences requiring no imputation, while the spatial block validation uses all 20,170 pixels with same-year nearest-neighbour spectral imputation applied before partitioning.
The authors note that one possible explanation for the sequence models' better average performance is that they capture ordered intra-seasonal spectral transitions and are more strongly regularized, but this mechanism was not directly isolated. Covariate shift between blocks (KS 0.11–0.27) may reflect both local soil and topographic variation and unrecorded preceding-crop effects. The T-CGAN gain for CNN-A-GRU on spatial block R² was not significant across the 5 independent fold means (p=0.112) and was significant only across the 25 fold-seed repeats (p=0.0090). Errors remain elevated in the 2–6 t/ha transition bins.
