Predicting Within-Field Yield Variation: an Evaluation Between UAV and Satellite Platforms

Abstract

This study explores the use of multispectral images from both Unmanned Aerial Vehicles (UAVs) and satellite platforms to build prediction models for within-field yield variations.

Satellite images from the early stages and UAV images from the middle and late stages of crop growth were collected from two spring wheat fields in Norway. From the images, reflectance values were extracted. Various features, such as vegetation indices, simple ratios, and a temporal Normalized Difference Vegetation Index (NDVI), were added to the datasets to assess their impacts on model performance. Harvesters from the fields provided detailed yield data.

The analysis includes regression and classification techniques, including random forest models, multilayer perceptron (MLP), and 1-dimensional Convolutional Neural Network (CNN). Correlation analysis of the data features against grain yield indicates strong correlations between the temporal NDVI and yield for both platforms, ranging from 0.18 to 0.47.

The findings from the model analyses reveal that random forest models integrating UAV reflectance data with additional non-linear features in simple ratios and vegetation indices have increased prediction scores compared to alternative models. These models outperformed the deep learning models in this context. The best-performing regression model had an R2 score of 0.638 and an MSE score of 0.292, and the best-performing classification model achieved an accuracy of 59.4%. These models correspond to the crops’ filling stage and arise from datasets with reflectance values and added vegetation indices. Models only utilizing reflectance values achieved an R2 score of 0.588, MSE score of 0.332, and accuracy of 56.5%.

Results from the satellite classification models indicate that augmenting the datasets to include non-linear features (simple ratios or vegetation indices) occasionally decreased the models’ performance.