Accurate production estimates, months before the harvest, are crucial for all parts of the food supply chain, from farmers to governments. While methods have been developed to use satellite data to monitor crop development and production, they typically rely on official crop statistics or ground-based data, limiting their application to the regions where they were calibrated. To address this issue, a new method called VeRsatile Crop Yield Estimator (VeRCYe) has been developed to estimate wheat yield at the pixel and field levels using satellite data and process-based crop models. The method uses the Leaf Area Index (LAI) as the linking variable between remotely sensed data and APSIM crop model simulations. In this process, the sowing dates of each field were detected (RMSE = 2.6 days) using PlanetScope imagery, with PlanetScope and Sentinel-2 data fused into a daily 3 m LAI dataset, enabling VeRCYe to overcome the traditional trade-off between satellite data that has either high temporal or high spatial resolution. The method was evaluated using 27 wheat fields across the Australian wheatbelt, covering a wide range of pedo-climatic conditions and farm management practices across three growing seasons. VeRCYe accurately estimated field-scale yield (R2 = 0.88, RMSE = 757 kg/ha) and produced 3 m pixel size yield maps (R2 = 0.32, RMSE = 1213 kg/ha). The method can potentially forecast the final yield (R2 = 0.78–0.88) about 2 months before the harvest. Finally, the harvest dates of each field were detected from space (RMSE = 2.7 days), indicating when and where the estimated yield would be available to be traded in the market. VeRCYe can estimate yield without ground calibration, be applied to other crop types, and used with any remotely sensed LAI information. This model provides insights into yield variability from pixel to regional scales, enriching our understanding of agricultural productivity.

Wheat accounts for more than 50% of Australia’s total grain production. The capability to generate accurate in-season yield predictions is important across all components of the agricultural value chain. The literature on wheat yield prediction has motivated the need for more novel works evaluating machine learning techniques such as random forests (RF) at multiple scales. This research applied a Random Forest Regression (RFR) technique to build regional and local-scale yield prediction models at the pixel level for three southeast Australian wheat-growing paddocks, each located in Victoria (VIC), New South Wales (NSW) and South Australia (SA) using 2018 yield maps from data supplied by collaborating farmers. Time-series Normalized Difference Vegetation Index (NDVI) data derived from Planet’s high spatio-temporal resolution imagery, meteorological variables and yield data were used to train, test and validate the models at pixel level using Python libraries for (a) regional-scale three-paddock composite and (b) individual paddocks. The composite region-wide RF model prediction for the three paddocks performed well (R2 = 0.86, RMSE = 0.18 t ha−1). RF models for individual paddocks in VIC (R2 = 0.89, RMSE = 0.15 t ha−1) and NSW (R2 = 0.87, RMSE = 0.07 t ha−1) performed well, but moderate performance was seen for SA (R2 = 0.45, RMSE = 0.25 t ha−1). Generally, high values were underpredicted and low values overpredicted. This study demonstrated the feasibility of applying RF modeling on satellite imagery and yielded ‘big data’ for regional as well as local-scale yield prediction.

We compared Support Vector Machine (SVM) and Random Forest (RF) machine learning approaches with the widely used Jarvis-type phenomenological model for predicting stomatal conductance (g(s)) in wheat (Triticum aestivum L.) using historical measurements collected in the Australian Grains Free-Air CO2 Enrichment (AGFACE) facility. The machine learning-based methods produced greater accuracy than the Jarvis-type model in predicting gs from leaf age, atmospheric [CO2], photosynthetically active radiation, vapour pressure deficit, temperature, time of day, and soil water availability (i.e. phenological and environmental variables determining gs). The R-2 was 0.76 for the Jarvis-type but 0.92 for SVM and 0.97 for RF machine learning-based models, with a calculated RMSE of 0.292 mol m(-2) s(-1) in the Jarvis-type compared to 0.129 mol m(-2) s(-1) in SVM and 0.081 mol m(-2) s(-1) in RF. The machine learning models, however, needed large datasets for training to achieve statistical significance, and do not offer the same opportunity to provide physiological insights through a statistically testable hypothesis. These results show that using the machine-learning based methods can achieve high prediction accuracy of gs that is especially important when incorporated into larger models, but their ability to extrapolate beyond observed data ranges will need to be assessed before they could be considered in place of the physical model.

Foresight of grain yields prior to harvest would be empowering for many stakeholders along the supply chain from farmers through to bulk handlers, banks and insurance companies. Estimating Australian grain production ahead of harvest is difficult for many reasons including the highly variable year to year rainfall. The rainfall in the final months prior to harvest, can be crucial to final harvest totals. Here we explore the importance of rainfall from September 1, which broadly corresponds to the close of the top-dressing fertilizer application window, for the remaining cropping season in determining final yield. This is assessed via sensitivity analysis of water-limited wheat potential yield totals from historical climate in the APSIM crop model. At locations where the rainfall influences wheat yield, we compare three methods to forecast wheat yields that differ based on the climate data input: 1) climatology approach, which uses 30 years of observed climate data, 2) analogue climatology, which uses information from climate drivers (El-Nino Southern Oscillation and Indian Ocean Dipole) to create analogue years; and 3) dynamical climate forecasts from a general circulation model (ACCESS-S). We find that potential yields strongly depend on in-season plant available water (PAW) where years with high PAW are unaffected by the late season rainfall. Predicting the potential yield from analogue climatology (climate drivers) had the greatest skill, with smallest Root Mean Squared Error of 0.45 t/ha. This approach ranked first for 42% of the study locations compared to the climatology and ACCESS-S forecasting methods. This knowledge can help inform decision makers about the need to incorporate seasonal climate forecasts and the most appropriate climate forecasting method.

Reducing the number of tillers per plant using a tiller inhibition (tin) gene has been considered as an important trait for wheat production in dryland environments. We used a spatial analysis approach with a daily time-step coupled radiation and transpiration efficiency model to simulate the impact of the reduced-tillering trait on wheat yield under different climate change scenarios across Australia's arable land. Our results show a small but consistent yield advantage of the reduced-tillering trait in the most water-limited environments both under current and likely future conditions. Our climate scenarios show that whilst elevated [CO2] (e[CO2]) alone might limit the area where the reduced-tillering trait is advantageous, the most likely climate scenario of e[CO2] combined with increased temperature and reduced rainfall consistently increased the area where restricted tillering has an advantage. Whilst long-term average yield advantages were small (ranged from 31 to 51 kg ha(-1) year(-1)), across large dryland areas the value is large (potential cost-benefits ranged from Australian dollar 23 to 60 MIL/year). It seems therefore worthwhile to further explore this reduced-tillering trait in relation to a range of different environments and climates, because its benefits are likely to grow in future dry environments where wheat is grown around the world.

The onus for monitoring crop growth from space is its ability to be applied anytime and anywhere, to produce crop yield estimates that are consistent at both the subfield scale for farming management strategies and the country level for national crop yield assessment. Historically, the requirements for satellites to successfully monitor crop growth and yield differed depending on the extent of the area being monitored. Diverging imaging capabilities can be reconciled by blending images from high-temporal-frequency (HTF) and high-spatial-resolution (HSR) sensors to produce images that possess both HTF and HSR characteristics across large areas. We evaluated the relative performance of Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat, and blended imagery for crop yield estimates (2009–2015) using a carbon-turnover yield model deployed across the Australian cropping area. Based on the fraction of missing Landsat observations, we further developed a parsimonious framework to inform when and where blending is beneficial for nationwide crop yield prediction at a finer scale (i.e., the 25-m pixel resolution). Landsat provided the best yield predictions when no observations were missing, which occurred in 17% of the cropping area of Australia. Blending was preferred when <42% of Landsat observations were missing, which occurred in 33% of the cropping area of Australia. MODIS produced a lower prediction error when ≥42% of the Landsat images were missing (~50% of the cropping area). By identifying when and where blending outperforms predictions from either Landsat or MODIS, the proposed framework enables more accurate monitoring of biophysical processes and yields, while keeping computational costs low.

Empirical yield estimation from satellite data has long lacked suitable combinations of spatial and temporal resolutions. Consequently, the selection of metrics, i.e., temporal descriptors that predict grain yield, has likely been driven by practicality and data availability rather than by systematic targetting of critically sensitive periods as suggested by knowledge of crop physiology. The current trend towards hyper-temporal data raises two questions: How does temporality affect the accuracy of empirical models? Which metrics achieve optimal performance? We followed an in silico approach based on crop modelling which can generate any observation frequency, explore a range of growing conditions and reduce the cost of measuring yields in situ. We simulated wheat crops across Australia and regressed six types of metrics derived from the resulting time series of Leaf Area Index (LAI) against wheat yields. Empirical models using advanced LAI metrics achieved national relevance and, contrary to simple metrics, did not benefit from the addition of weather information. This suggests that they already integrate most climatic effects on yield. Simple metrics remained the best choice when LAI data are sparse. As we progress into a data-rich era, our results support a shift towards metrics that truly harness the temporal dimension of LAI data.

Most cropping systems around the world are organised around few dominant crops and a larger number of less frequent crops. While rare and infrequent crops occupy a small share of the cropped area, they produce ecological benefits on farmland, contribute to sustainability and help provide food and nutritional security. However, data about their location and extent derived from satellite imagery generally lack accuracy, largely owing to the class imbalance problem. Class imbalance occurs when only few instances of some classes are available for training classifiers, and leads to large error rates of the infrequent classes. In this study, we assessed the magnitude of the class imbalance problem in crop classification and evaluated balancing methods to combat it by creating synthetic minority observations or by removing majority observations. To that aim, we generated 18 unbalanced data sets from Sentinel-2 time series and crop type observations in Victoria, Australia. These data sets covered a wide range of complexity, number of classes, number of samples per class and spectral separability which enabled us to gather evidence about the benefits and drawbacks of balancing methods in various settings. Classification accuracy was assessed with two metrics: the Overall Accuracy (OA), which gives more weight to majority classes, and the G-Mean accuracy (GM), which is more sensitive to minority classes. Results showed that class imbalance explained near 40% of the accuracy variability. We found that balancing methods boosted GM by 0.01–0.54 but no single best solution emerged. The price for increasing the accuracy of minority classes was a drop in OA of a magnitude that was problem- and method-specific. We thus applied an algorithm selection method called the F-race to identify optimal balancing methods in a computationally economic fashion. Optimal balancing methods lead to maximum gain in GM and minimum loss in OA. We demonstrated that this approach either successfully identified optimal balancing methods or ones that were not significantly sub-optimal, while reducing the computational cost by up to 60%. It can readily be incorporated to operational crop classification systems with little disruption to the existing processing chains. This contribution paves the way for achieving a more comprehensive and detailed view of crop distribution and cropping sequences.

There is considerable demand for nationwide grain yield estimation during the cropping season by growers, grain marketers, grain handlers, agricultural businesses, and market brokers. In this paper, we developed a semi-empirical model (Crop-SI) to estimate the yield of the three major crops in the dryland Australian wheatbelt by combining a radiation use efficiency approach with meteorology driven Stress Indices (SI) at critical crop growth stages (e.g., anthesis and grain filling). These crop-specific SI (e.g., drought, heat and cold stress) help explain the impact of high spatial agro-environmental heterogeneity, which lead to substantial improvement in grain yield prediction. Crop-SI explains 87%, 69% and 83% of the observed field-scale grain yield variability with root mean square error of ~0.4, 0.4 and 0.5 t/ha for canola, wheat, and barley, respectively. At the pixel-level, Crop-SI reduces the relative error in grain yield estimation to 34%, 25%, and 20% for canola, wheat, barley, respectively, compared to two benchmark models. By incorporating water- and temperature-driven stresses, Crop-SI's predictive skill in highly variable environments is enhanced. As such, it paves the way for the next generation of agricultural systems models, knowledge products and decision support tools that need to operate at various scales.

Large-area empirical yield estimation from space has long been constrained by the lack of suitable combinations of spatial and temporal resolutions of satellite image time series. Consequently, it is likely that the selection of metrics, i.e., time series descriptors used to predict grain yield, was driven by practicality and data availability rather than by systematic targetting of critically sensitive periods suggested by knowledge of crop physiology. The availability of data from new satellites is driving a trend towards increased spatial and temporal resolutions. This raises the question of how temporality affects performance and the choice of metric required to achieve optimal performance. To answer this question, we followed an in silico approach and generated daily time series of Leaf Area Index (LAI) using a crop growth model. In comparison with currently limited observational data, this approach can generate any observation frequency, explore a wider range of growing conditions and drastically reduce the cost of measuring yields in situ. Wheat crops were thus simulated under nine management strategies for 30 years at 50 locations representative of the Australian grain zone. Six types of LAI metrics were extracted to predict grain yields. Between 30% to 80% of the yield variability could be explained depending on the metric. The high accuracy achieved by advanced metrics (Senescence fit and Fourier decomposition) is remarkable since it is obtained from linear models calibrated for the whole of the grain zone, contrasting seasons, and multiple cultivars. Adding weather variables doubled the R of models based on simple metrics (peak LAI and Early/Late window) but did not significantly improve models based on advanced metrics. This constitutes strong evidence that metrics intensively exploiting the temporal dimension of LAI already integrate the information content of weather variables. Simple metrics remain nonetheless the best choice when data are sparse. As we progress in a data-rich era, these results provide evidence for a shift towards the inclusion of metrics that truly harness the temporal dimension of LAI time series.

There has been increasing recognition for the need to reform doctoral training practices to foster students' personal epistemology. This study describes the design and evaluation of a learning experience designed to help students understand the scholarly publication process. Firstly, this study discusses the design of the learning experience, describing the collaborative process of writing an interdisciplinary publication using both online and face-to-face learning. Secondly, this study evaluates the effectiveness of the learning experience by examining students' reflections. We show that participation in the learning experience helped students to develop their academic writing proficiency, collaboration and teamwork, intercultural competence, and ability to engage in reflective practice. Importantly, we show that each student also created more individualised knowledge, gaining insight into how they and others think. This study, therefore, demonstrates that personal epistemology can be fostered through collaboration in a doctoral writing group context.

Accurate description of forest surface fuel load is important for understanding bushfire behaviour and suppression difficulties, predicting ongoing fires for operational activities, assessing potential fire hazards and assisting in fuel hazard-reduction burns to reduce fire risks to the community and the environment. Bushfire related studies and current operational activities have a common challenge in quantifying fuels, since the fuel load varies across the landscape. This paper developed a predictive model that efficiently and accurately estimates quantities of surface fuel in Australian southeast Eucalypt forests. Model coefficients were determined through a three-step process that attempts to evaluate how the spatial variation in surface fuel load relates to litter-bed depth, fuel characteristics, topography and previous fire disturbance. First, the forest surface fuel depth-to-load relationship was established; second, key quantitative variables of environmental factors were added; and third, important qualitative variables of fuel characteristics were included. The verification of model prediction was conducted through leave-one-out cross-validation (CV). Light Detection and Ranging was used to quantify forest structural characteristics and terrain features. The calibrated model had a R-2 of 0.89 (RMSE = 20.7 g) and performed better than the currently used surface fuel load models, including McArthur's (R-2 = 0.61 and RMSE = 39.6 g) and Gilroy and Tran's (R-2 = 0.69 and RMSE = 36.5 g) models. This study describes a novel approach to forest surface fuel load modelling using forest characteristics and environmental factors derived from LiDAR data through statistical analysis. The model established in this study can be used as an efficient approach to assist in forest fuel management and fire related operational activities. (C) 2017 Elsevier Ltd. All rights reserved.

Accurately describing forest surface fuel load is significant for understanding bushfire behaviour and suppression difficulties, predicting ongoing fires for operational activities, as well as assessing potential fire hazards. In this study, the Light Detection and Ranging (LiDAR) data was used to estimate surface fuel load, due to its ability to provide threedimensional information to quantify forest structural characteristics with high spatial accuracies. Firstly, the multilayered eucalypt forest vegetation was stratified by identifying the cut point of the mixture distribution of LiDAR point density through a non-parametric fitting strategy as well as derivative functions. Secondly, the LiDAR indices of heights, intensity, topography, and canopy density were extracted. Thirdly, these LiDAR indices, forest type and previous fire disturbances were then used to develop two predictive models to estimate surface fuel load through multiple regression analysis. Model 1 was developed based on LiDAR indices, which produced a R-2 value of 0.63. Model 2 (R-2 = 0.8) was derived from LiDAR indices, forest type and previous fire disturbances. The accurate and consistent spatial variation in surface fuel load derived from both models could be used to assist fire authorities in guiding fire hazard-reduction burns and fire suppressions in the Upper Yarra Reservoir area, Victoria, Australia.

Fuel structural characteristics affect fire behavior including fire intensity, spread rate, flame structure, and duration, therefore, quantifying forest fuel structure has significance in understanding fire behavior as well as providing information for fire management activities (e.g., planned burns, suppression, fuel hazard assessment, and fuel treatment). This paper presents a method of forest fuel strata classification with an integration between terrestrial light detection and ranging (LiDAR) data and geographic information system for automatically assessing forest fuel structural characteristics (e.g., fuel horizontal continuity and vertical arrangement). The accuracy of fuel description derived from terrestrial LiDAR scanning (TLS) data was assessed by field measured surface fuel depth and fuel percentage covers at distinct vertical layers. The comparison of TLS-derived depth and percentage cover at surface fuel layer with the field measurements produced root mean square error values of 1.1 cm and 5.4%, respectively. TLS-derived percentage cover explained 92% of the variation in percentage cover at all fuel layers of the entire dataset. The outcome indicated TLS-derived fuel characteristics are strongly consistent with field measured values. TLS can be used to efficiently and consistently classify forest vertical layers to provide more precise information for forest fuel hazard assessment and surface fuel load estimation in order to assist forest fuels management and fire-related operational activities. It can also be beneficial for mapping forest habitat, wildlife conservation, and ecosystem management. (C 2016 Society of Photo-Optical Instrumentation Engineers (SPIE)

针对多云雾地区高时空分辨率数据缺乏现状,提出了一套区域尺度高时空分辨率植被覆盖度数据构建方法.首先,通过时空适应反射率融合模型(STARFM)有效地将TM的较高空间分辨率与MODIS的高时间分辨率融合在一起,构建了研究区植被生长峰值阶段的NDVI数据;然后,以植被生长峰值阶段的NDVI为输入,基于地表覆被类型,综合应用等密度和非密度亚像元模型对研究区的植被覆盖度进行估算.结果表明:①即使数据源存在大量的云雾,且存在一定的时相差异,研究区植被覆盖度的估算结果过渡自然,不存在明显的不接边效应;②以植被生长峰值阶段的NDVI数据为输入进行植被覆盖度估算,有效拉开了同一地表覆被类型不同覆盖度像元的NDVI梯度,提高了亚像元估算模型对输入数据的抗扰动性;③基于地表覆被类型,应用亚像元混合模型,能够提高植被覆盖度的估算精度.经野外实测数据验证,总体约85％的估算精度表明,针对高时空分辨率遥感数据缺乏的多云雾区域,本研究提出的方法能够实现区域尺度植被覆盖度数据的构建.

Background: A number of data sources currently exist that can provide information on forest plantations at a range of scales over an entire rotation cycle. In particular, LiDAR is quickly becoming the technology of choice for harvest planning and providing local-scale estimates of forest structure. Its application is still limited as repeat annual acquisition at this scale is generally cost prohibitive. Development of temporally updateable models that can accurately project important metrics such as tree height between LiDAR acquisitions would be of considerable use to resource managers. The objective of this research was to develop models of Pinus radiata height using GIS spatial data supplemented with RapidEye satellite imagery.Methods: Multiple regression models were constructed to describe maximum canopy height (H-m) derived from LiDAR at two relatively distant study sites located in Kaingaroa and Tairua forests. A randomised selection of 300 m(2) circular plots was made at both sites and average values of Hm within these plots were used for the modelling. Sources of information used for predicting H-m included stand age and spatial information describing environmental variables and stand productivity. This information was supplemented with spectra and vegetation ratios derived from high resolution RapidEye satellite imagery.Results: The most robust models of H-m that were developed for both sites included a combination of the crop age obtained from the stand GIS, Site Index (obtained from a GIS surface) and the red-edge vegetation ratio (REVI) The final models of Hm had respective R-2 of 0.99 and 0.94 for the Kaingaroa and Tairua sites. At both sites, stand age was the strongest predictor of Hm. However, the inclusion of REVI from high resolution imagery did add an updatable temporal dimension to the model. Changes in REVI are sensitive to the impacts of abiotic and biotic factors that are not captured by stand age and Site Index.Conclusion: Applied operationally, this model can be used in a GIS environment to estimate tree height and identify areas of anomalous growth or disturbance caused by wind, snow, fire or disease.

This article presents an integrated GIS tool for automatic forest inventory of Pinus radiata plantations from light detection and ranging (LiDAR) data. Built as a set of tools running in the desktop GIS software package ArcGIS, it integrates spatial analysis, LiDAR data analysis and image segmentation techniques as well as empirical tree models to support forest inventories of Pinus radiata on an individual-tree basis. The integrated GIS tool allows users to define or select plots to extract LiDAR data for forest inventory, build and process canopy height models (CHMs) from the extracted LiDAR data through surface modelling, delineate individual trees on the CHMs by applying the marker-controlled watershed segmentation technique, and to derive forest inventory estimates based on the CHMs and identified individual trees through spatial analysis and tree modelling using the empirical models. It takes advantage of combining GIS and LiDAR to automatically conduct a forest inventory, build and manage a forest inventory database and to spatially and statistically summarise and visualise the inventory data. Although developed for forest inventories of Pinus radiata plantations in Victoria, Australia, the integrated tool can be customised for forest inventories of Pinus radiata plantations in other regions and for other types of plantations by incorporating empirical tree models built and calibrated for those regions and those species of trees.

The aim of site quality assessment of Pinus radiata plantations is to determine the quality and productivity of the growing stock at different sites. It provides a useful indication of the site productivity to assist in the allocation of optimum thinning and fertiliser regimes and the scheduling of silvicultural operations. The predominant stand height (PDH) at a specific reference age, also known as site index (SI), is often used for site quality assessment of Pinus radiata plantations in Australia, as it is closely correlated with site productivity. However, measuring PDH in the field can be a time- and resource-consuming task. This paper proposes the use of light detection and ranging (LiDAR) data to estimate PDH for assessing the site quality of Pinus radiata. LiDAR provides highly accurate digital elevation and surface data that can be used to build a canopy height model (CHM). In this study, the state-of-the-art image segmentation technique, marker-controlled watershed segmentation, was employed for identifying locations of individual trees and estimating their heights from a CHM. Using an empirically derived SI equation, PDHs with reference age 11 years (SI11) were estimated from the tallest trees identified in each forest stand, and were then used to determine the site quality class for each stand. The comparison of LiDAR-derived tree heights with field measurements produced an RMSE value of 0.42 m. The maximum horizontal distance between the field-measured locations of individual trees and the LiDAR-detected locations of their treetops was 1.87 m. Site quality classification was conducted in terms of 0.05 ha gridded plots, which revealed more detailed spatial variations of site quality across the study area than classification based on management plots. The study demonstrated that LiDAR provides an effective and accurate method for site quality classification of Pinus radiata.