A knowledge-guided machine learning framework for cross-scale wheat harvest monitoring via sample augmentation

Accurate monitoring of wheat harvest is crucial for precision agriculture and ensuring food security. However, rapid changes in land surface composition during the harvest period in intensive agricultural regions make it difficult to obtain sufficiently high-confidence ground samples, limiting the performance and generalization of data-driven remote sensing methods. Therefore, this study proposes a Knowledge-Guided Machine Learning (KGML) framework that integrates multi-satellite Earth observation data (PlanetScope, Sentinel-2, and MODIS) to monitor harvest from the field to regional scales. Ground data were collected using vehicle-mounted cameras and smartphones during the 2023 and 2024 wheat harvest periods. The results showed that combining spectral knowledge rules with a Random Forest model (regional accuracy >0.80) generated numerous high-confidence augmented samples from PlanetScope imagery. The augmented dataset was used to train a Hybrid CNN-Transformer-LSTM (HCTL) model with two pathways: Sentinel-2 classification for field-level harvest mapping (overall accuracy = 0.93) and MODIS regression for sub-pixel harvest fraction estimation, which showed high agreement with PlanetScope-derived harvest fractions (R 2 = 0.97, RMSE = 0.07, rRMSE = 0.15). The harvest dates derived from the MODIS harvest fraction time series showed high consistency with field observations (R 2 = 0.82, RMSE = 1.30 days). This framework provides an effective solution for wheat harvest monitoring by bridging the gap between limited ground-truth data and multi-scale satellite observations, thereby supporting food security assessments and informed agricultural management decisions.