Modeling Mid-Season Rice Nitrogen Uptake Using Multispectral Satellite Data
Mid-season nitrogen (N) application in rice crops can maximize yield and profitability. This requires accurate and efficient methods of determining rice N uptake in order to prescribe optimal N amounts for topdressing. This study aims to determine the accuracy of using remotely sensed multispectral data from satellites to predict N uptake of rice at the panicle initiation (PI) growth stage, with a view to providing optimum variable-rate N topdressing prescriptions without needing physical sampling. Field experiments over 4 years, 4–6 N rates, 4 varieties and 2 sites were conducted, with at least 3 replicates of each plot. One WorldView satellite image for each year was acquired, close to the date of PI. Numerous single- and multi-variable models were investigated. Among single-variable models, the square of the NDRE vegetation index was shown to be a good predictor of N uptake (R² = 0.75, RMSE = 22.8 kg/ha for data pooled from all years and experiments). For multi-variable models, Lasso regularization was used to ensure an interpretable and compact model was chosen and to avoid over fitting. Combinations of remotely sensed reflectances and spectral indexes as well as variety, climate and management data as input variables for model training achieved R² < 0.9 and RMSE < 15 kg/ha for the pooled data set. The ability of remotely sensed data to predict N uptake in new seasons where no physical sample data has yet been obtained was tested. A methodology to extract models that generalize well to new seasons was developed, avoiding model overfitting. Lasso regularization selected four or less input variables, and yielded R² of better than 0.67 and RMSE better than 27.4 kg/ha over four test seasons that weren’t used to train the models.
The influence of nitrogen and variety on rice grain moisture content dry-down
Rice field management around maturity and harvest are some of the most difficult decisions growers face. Field drainage and harvest timing affect quality, yield, and post-harvest drying costs. These decisions are informed by grain moisture content (MC). Over three years, three sites and three varieties, we studied the field dry-down rate and time to optimal harvest MC. We showed that field-specific parameters significantly affected these characteristics, including rice variety, Nitrogen applied (NA), mid-season N uptake (NU) and dry matter (DM). Increased N and DM is associated with increased MC and thus delays time to harvest. We developed models based on linear regression and nonlinear machine learning (ML) algorithms, including parameters describing these field-specific conditions. Cross validation across the three years provided a realistic expectation of model prediction errors. A linear model with the addition of nonlinear predictors achieved competitive performance compared with more complex and less interpretable ML models. When MC was modeled as a function of days since heading, similar or better accuracy was achieved to using accumulated weather parameters. Moisture content was predicted with mean absolute error of 2.1 %. The predicted time from heading to harvest MC was improved by the inclusion of field-specific parameters (N and variety) from mean absolute error of 6.8 days to 5.7 days. The final linear regression model explained 80 % of the moisture variability in the dataset, and provided estimates of dry-down rates, moisture as a function of time, and time to reach harvest moisture. This study shows the importance of including field-specific parameters when estimating of rice harvest timing, and provides methods to model these effects.
Rice nitrogen status detection using commercial-scale imagery
Determining the mid-season nitrogen status of rice is important for precision application of fertilizer to optimize productivity. While there has been much research aimed at developing remote-sensing-based models to predict the nitrogen status of rice, this has been predominantly limited to scientific small plot trials, relying on experts performing radiometric calibrations, encompassing limited cultivars, seasons and locations, and uniform management practices. As such, there has been little testing of models at commercial scale, against the range of conditions encountered across entire growing regions. To fill this gap, this work brings together four years of data, from both experimental replicated plot trials (38 datasets with 1734 observations) and commercial farms (12 datasets with 106 observations). Using commercial scale imagery acquired from airplanes, a number of nitrogen uptake modeling methodologies were evaluated. Universal single vegetation index based linear regression models had prediction root mean squared error (RMSE) of more than 45 kg/ha when tested at the 12 commercial sites. Machine learning models using multiple remote sensing features were able to improve predictions somewhat (RMSE > 30 kg/ha). Practically useful accuracies were achieved after using three local field samples to calibrate models to each field image. The prediction RMSE using this methodology was 22.9 kg/ha, or 19.4%. This approach enables provision of optimal variable-rate mid-season rice fertilizer prescriptions to growers, while motivating continued research towards development of methods that reduce requirement of local sampling.
Analysis and forecasting of Australian rice yield using phenology-based aggregation of satellite and weather data
Rice yield depends on factors including variety, weather, field management, nutrient and water availability. We analyzed important drivers of yield variability at the field scale, and developed yield forecast models for crops in the temperate irrigated rice growing region of Australia. We fused a time-series of Sentinel1 and Sentinel-2 satellite remote sensing imagery, spatial weather data and field management information. Rice phenology was predicted using previously reported models. Higher yields were associated with early flowering, higher chlorophyll indices and higher temperatures around flowering. Successive rice cropping in the same field was associated with lower yield (p<0.001). After running a series of leave-one-year-out cross validation experiments, final models were trained using 2018–2022 data, and were applied to predicting the yield of 1580 fields (43,700 hectares) from an independent season with challenging conditions (2023). Models which aggregated remote sensing and weather time-series data to phenological periods provided more accurate predictions than models that aggregated these predictors to calendar periods. The accuracy of forecast models improved as the growing season progressed, reaching RMSE=1.6 t/ha and Lin’s concordance correlation coefficient (LCCC) of 0.67 30 days after flowering at the field level. Explainability was provided using the SHAP method, revealing the likely drivers of yield variability overall, and of individual fields.
Impact of UAV Time-of-Flight on Rice Nitrogen Uptake Models
This work examines the impact of image acquisition time on rice paddy remote sensing data in Yanco, NSW, Australia. The normalized difference red edge (NDRE) is an important vegetation index for rice paddy N status. Models of midseason N uptake were extracted using physical samples and NDRE data taken at multiple times of day from a UAV with multi-spectral camera. The best prediction error was 15.8 kg/ha. With a model extracted at 13:00 on 1 January, model predictions from 27 December to 11 January, at image times from 11:00 to 16:00, N uptake prediction errors were less than 28.5 kg/ha. Predictions from images acquired at times very different from the one used to extract the model are more erroneous.
Rice ponding date detection in Australia using Sentinel-2 and Planet Fusion imagery
Rice is unique, in that yields are maximized when it is grown under ponded (or flooded) conditions. This however has implications for water use (an important consideration in water-scarce environments) and green-house gas emissions. This work aimed to provide precise predictions of the date when irrigated rice fields were ponded, on a per-field basis. Models were developed using Sentinel-2 data (with the advantage of inclusion of water-sensitive shortwave infrared bands) and Planet Fusion data (which provides daily, temporally consistent, cross-calibrated, gap-free data). Models were trained with data from both commercial farms and research sites in New South Wales, Australia, and over four growing seasons (harvest in 2018–2021). Predictions were tested on the 2022 harvest season, which included a variety of sowing and water management strategies. A time-series method was developed to provide models with features including satellite observations from before and after the date being classified (as ponded or non-ponded). Logistic regression models using time-series features produced mean absolute errors for ponding date prediction of 4.9 days using Sentinel-2 data, and 4.3 days using Planet Fusion data. The temporal frequency of the Planet Fusion data compensated for the lack of spectral bands relative to Sentinel-2.
Predicting rice phenology and optimal sowing dates in temperate regions using machine learning
Crop phenology modeling often involves determining variety-specific growing degree day thresholds, or parameterizing mechanistic crop models. In this work, we used machine learning methods to develop models that provide daily predictions of the probability that rice (Oryza sativa) crops had reached the panicle initiation and flowering growth stages. These per-date classifications were summarized into perpaddock growth stage transition dates, which were then compared with field-sampled reference data, encompassing 15 rice varieties, 10 years, and 380 sites. Leave-oneyear-out cross validation was used to provide realistic estimates of model errors. Compared with more complex and computationally intensive algorithms, logistic regression produced competitive results (mean cross-season validation RMSE 3.9 and 5.2 days for panicle initiation and flowering, respectively). Logistic regression had additional advantages: providing confidence of growth stage predictions at each date (as it is a probabilistic algorithm), and straightforward explainability (as model parameters directly indicated how the various input variables contributed to growth stage predictions). Input variables included accumulated weather, rice variety, and sowing methods. The models were applied to forecasting phenology transition dates of the rice crops planted throughout the Murray and Murrumbidgee valleys. In addition, recommendations for optimal sowing dates were developed, using simulations involving more than 40 years of weather data, with the goal of minimizing the risk of cold-temperatures during the microspore growth phase, which can severely degrade yield in temperate rice growing regions.