Real-time predictive modelling of rice crops to optimize field management

The sustainability and productivity of temperate rice farming can be enhanced by optimizing timing of field management, such as sowing, fertilization, water application and draining, and harvest. We are developing tools to support this aim, leveraging many years of Australian rice agronomic field trials and industry data. Specifically, we are bringing together field-level data (e.g. phenology observations and yield) and spatio-temporal data (satellite imagery and weather) to train machine learning models. These models are deployed in a cloud-based software system, that provides in-season predictions to growers and industry in near real-time. In this paper, we describe a subset of the project outputs, including in-season rice maps, growth curve monitoring, water application date detection and phenology prediction. To produce the rice maps, we use a time-series of Sentinel-1 and 2 (radar and optical) satellite imagery together with random forest classifications on a per-pixel basis (10-meter resolution). These are then spatially clustered and vectorized to generate shapefiles of all rice fields. In 2023, we generated biweekly maps, which achieved accuracy metric (F1score) of 80% at 1 January (before panicle initiation), and 91% at 15 February (flowering). We derive smoothed time-series vegetation indicators for all paddocks from satellite data, which are benchmarked against 5 previous years of similar data across all rice fields in the growing region. This facilitates understanding of the timing and rate of biomass accumulation of crops, which can be used to guide mid-season fertilizer application decisions. We detect the start of permanent ponding for each paddock using a logistic regression classification model, based on Sentinel-2 time series. The mean absolute error (MAE) was 3.7 days in 2023. Logistic regression models for panicle initiation and flowering date used variety, accumulated weather and sowing method as inputs. 2023 model prediction MAEs were 2.7 and 5.3 days, respectively. Interactive web dashboards were provided to growers, including the above predictions among others, and were updated daily with the latest data. We anticipate that this information will aid decisions leading to improved productivity in terms of yield and water efficiency.