Answers are grounded in this paper and cite their sources.
Reliable crop-stress diagnosis and yield forecasting require observations that capture both short-term canopy responses and seasonal crop development. Proximal sensors resolve canopy structure and physiology, whereas satellite time series provide phenological context; their integration is complicated by unequal footprints, asynchronous acquisition, variable quality, and missing observations. We developed CausalScale-Former, a multimodal phenotyping framework that aligns proximal and satellite observations by footprint and crop stage, learns from masked observations, and weights modalities using recorded quality. Two TERRA-REF sorghum seasons were combined with Sentinel-2 time series, environmental records, and a weak crop prior from the USDA Cropland Data Layer. All records from each plot remained in one partition. Evaluation included bidirectional seasonal transfer, held-out genotype and irrigation groups, spatial blocks, and natural or simulated data gaps. Stress Macro-F1 was 0.842 in plot-grouped testing, 0.755 and 0.766 in the two seasonal-transfer tests, and 0.738 in the spatial-block analysis (Table 13). Event-level F1 was 0.716, 0.742, and 0.721 at 7, 10, and 14 days, respectively (Table 11). At the final pre-harvest cutoff, yield prediction achieved and RMSE of 0.82 t ha (Table 23, Panel A). Within this two-season, single-site experiment, scale- and phenology-aware multimodal phenotyping supported joint crop-stress diagnosis, early warning, and probabilistic yield prediction under incomplete sensing. External multi-site and multi-year validation is required before operational deployment. The causality-inspired regularizer does not identify causal effects.
No takes yet. Share an insight, caveat, or question.
Cui et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: