Corn grain and silage yield class prediction for zone delineation using high-resolution satellite imagery

CONTEXT: Reliable stability zone delineation requires considering spatial variability from each year and temporal variability across at least three years. Yet, in cases where farms lack either temporal or spatial records, leveraging remote sensing and machine learning to predict corn (Zea mays L.) yield classes should be studied. OBJECTIVE: This study aimed to assess the accuracy of corn yield class prediction models and their practical applications: (i) adding yield maps for years with missing yield monitor data from the same field; and (ii) expanding yield mapping to corn fields without yield monitor data on the same farm. The predicted yield class maps were used to delineate stability zone maps and compared against the zones delineated from yield monitor data. METHODS: Our dataset consisted of yield monitor data for three consecutive years of corn grain and silage fields, across three farms in New York, USA. For each farm, yield data were categorized into five yield classes, ranging from low- to high-yielding. The multispectral satellite imagery from PlanetScope's CubeSat, digital elevation model (DEM) data, and a landform map derived from the DEM were used as remotely sensed data layers. Five soil indices and seven vegetation indices (VI) were calculated from bare soil imagery (close to planting) and top-ofcanopy imagery (at peak growth). We employed three feature selection methods: Boruta algorithm, recursive feature elimination, and sequential feature selection. The Bayesian optimization approach was used for hyperparameter tuning of three machine learning (ML) classifiers: random forest, support vector machine, and logistic regression. The best ML model was used to test the following two scenarios of yield prediction: (i) Year -fit - training with two years of data and predicting yield class of the third year; and (ii) Farm -fit - training with one field and predicting the yield of another field within the same farm on the same year. RESULTS AND CONCLUSIONS: The results indicated that all feature selection methods consistently identified elevation, five VIs, and one soil index as important features. The random forest classifier model gave the best accuracy, with hyperparameters specific to each farm. The classification accuracies of Year -fit ranged from 64 to 80% for silage and 65 to 76% for grain. Farm -fit ranged from 66 to 82% for the silage and 77 to 82% for grain. The zone maps delineated from the predicted yield classes resulted in accuracies of 66 to 91% for Year -fit and 64 to 85% for Farm -fit. SIGNIFICANCE: These results underscore the signification of expanding yield class prediction to multiple years, offering a practical approach for enhancing zone delineation applications.