Spring wheat trait prediction using combined multi-environment, weather and multispectral timeseries UAV data

Pre-harvest yield forecast is an important factor in high-throughput plant phenotyping (HTPP). It also is critical for making policy decisions concerning food security. The developments in Unmanned Aerial Vehicle (UAV) technology have provided cost-effective way to collect data from fields. Spectral bands data captured by UAVs is being used for increasing productivity, speeding up the breeding process, and select plants for desirable traits. However, most of the studies are limited to a single field or season. This study explores the possibility of using the data from multiple seasons and environments, along with weather information, to predict grain yield (GY) and days to maturity (DM) of spring wheat cultivars. The inclusion of weather data and spectral indices can increase the prediction score for wheat traits prediction for multi-environment, multi-spectral, timeseries remote sensing data of spring wheat cultivars. To validate this hypothesis, couple of integration methods were used to aggregate the multi-spectral data captured on several dates to make them comparable and usable together in a machine learning model. The multi-spectral data was collected at two different environments in 2019 and 2020(at both Vollebekk, Ås; Norway and Hamar, Staur, Norway). Grain yield and days to maturity were the wheat traits tested with the proposed method. A range of 28 spectral indices and 80 variations of weather features were tested along with 5 spectral bands, using multiple machine learning models to assess their contribution in predicting grain yield and days to maturity of spring wheat cultivars. The data from multiple dates was separately integrated using Simpson’s rule and Composite Trapezoidal Rule and the results identified that the latter consistently delivers better prediction score. The results confirm the hypothesis that the addition of weather data in the feature set improves the prediction score for both target variables, compared to models without weather data. Spectral bands did not contribute much to the prediction score of the models, with spectral indices, along with weather features, usually being more important features contributing towards prediction score. The processed datasets with machine learning models were more accurate in predicting days to maturity than they were for grain yield prediction. Sequential feature selection was applied to address multicollinearity in the features set. A group of 11 features were identified that could predict grain yield with comparable prediction score instead of using the complete features set. For days to maturity prediction, a set of 23 features were selected which could be even further reduced to 10 features. The results of this study support the integration using composite trapezoidal rule, for data collected on multiple dates, and inclusion of weather data for developing wheat traits prediction models.