Monitoring every crop sequence across the Western Australian low rainfall zone

Key messages

• The productivity of paddock-scale crop rotations can be monitored across many farms with remote sensing of crop type and crop modelling with satellites.
• Over four years, rotations in the low rainfall zone of WA included wheat after cereal (41% of the area), cereal after oilseed (7.2% of the area), cereal after pulse or legume (5.7% of the area), cereal after pasture (37% of the area) and cereal after fallow (8.4% of the area)
• Compared to continuous cereal, legume crops improved the yield of the following wheat crop by ~150kg/ha. Wheat yields following canola and fallow were on average 180 and 170kg/ha higher than wheat following a cereal.  Wheat yields following a pasture were on average 60kg/ha higher than wheat following a cereal.
• Paddocks, farms and regions that either had lower or higher break crop effects than their neighbours were also identified.
• The analysis demonstrates how big data analytics can monitor crop management across WA, where consistently derived information can be provided to every farm and paddock in the wheatbelt.

Aims

Remote sensing and satellite imagery are increasingly being used to monitor crop production. These approaches have focussed on estimating crop yield and crop species but have not provided farmers with new information, or particularly novel management insights. Farmers know what crops are planted where, and many techniques exist to predict crop yield in a paddock. The real power of these digital approaches will be recognised when multiple pieces of information are combined. The aim in this project was to demonstrate how remote sensing of crop yield, and crop ID could be combined to create a useful management insight about crop rotation for farmers. We used this information to determine if farmers achieved similar “break crop effects” in cereals to what is widely reported by researchers.

Introduction

Crop production can now be monitored across the Australian landscape. Crop yields can be estimated with satellite imagery using the C-Crop model (Donohue et al 2018). Crop species can also be identified using satellite imagery and machine learning (Waldner et al 2019). To date, crop monitoring and yield evaluation has focused on capturing this information quickly and delivering this information to agribusiness to assist with logistics and marketing decisions. In contrast, farmers already know what crop is planted where, and may have other means of estimating actual yield and yield potential on their farm, through a variety of commercially available platforms, (e.g. yield monitors). Famers have used satellite imagery to assist, identify and manage within-paddock variability and correct various chemical and physical problems with soils. However, even these approaches do not link the observed variation to a management action. Rather, they identify a pattern in the landscape, and apply domain knowledge about the soil constraint or nutrient deficiency to prescribe a management action. Here we demonstrate an approach that focuses on assessing how effective crop rotation is in providing a break crop benefit to farmers across the low rainfall zone (LRZ) of Western Australia.

Method

Crop identification was conducted in 2016, 2017, 2018 and 2019. Optical satellite imagery captured throughout the growing season from June to October was extracted from the United States Geological Survey. Training data that identified which crops were growing where was captured via an annual field survey of the WA wheatbelt. Machine learning algorithms were used to characterise the agricultural land use across the WA wheatbelt. Land-uses included specifying the crop species, and identifying whether the paddock was in pasture or fallow. Details of the training data, image acquisition, crop areas and uncertainty estimates for the Crop ID classification are described by Waldner et al (2019). Data from all years were spatially joined in a geographic database to create a library of information about crop rotation, where the previous crop, as well as the current crop, were identified.

Crop yield was estimated with remote sensing using the C-Crop model (Donohue et al 2018). Crop yields for wheat, barley and canola were predicted with the model for paddocks identified as the particular crop species.

Both crop yield and crop ID are defined on a raster and needed to be clipped and aggregated to a particular field. Field boundaries for the entire WA wheatbelt were defined using the methods of Waldner and Diakogiannis (2020).

Weather data for each field was then extracted from the SILO database, and for each field, the growing season rainfall (April to October) was calculated.  Data about weather, crop yield, crop type and the previous crop were then extracted to explore the impact of crop rotation on wheat yield, across the LRZ of the WA wheatbelt.  Information from 43,481 fields were collated to explore the impact of crop rotation.

A simple grouped linear regression analysis was performed on estimated wheat yield. The previous crop, growing season rainfall and the interaction between the two main effects were fitted to explain wheat yield. That is, both the slope, in response to growing season rainfall, and the intercept, were allowed to vary with previous crop type. No other data were used in this relatively simple analysis.  The residuals (error) are also displayed graphically to determine if an individual field performed better or worse than the predicted effect from the statistical model.

Results

Crop rotation, growing season rainfall and their interaction accounted for 26% of the variation in predicted wheat yield from the C-Crop model. Growing season rainfall accounted for 23% of the variation, crop rotation accounted for 2.7%, while the interaction between the two main effects of growing season rainfall and crop rotation accounted for just 0.21% of the variation in predicted wheat yield. These results are presented in Figure 1. With 150mm of rainfall, wheat grown after a pulse and an oilseed yielded the highest. Small (~100kg/a) differences appeared between these two crop rotations and wheat crops grown after fallow, pasture or wheat. At 200mm of growing season rainfall, wheat after a fallow, pulse and oilseed yielded similarly, at around 1.5t/ha. Wheat yields after pasture were approximately 120kg/ha lower, and wheat yields after a cereal were 50kg/ha lower again.  At 250mm of rainfall, wheat after a fallow yielded more than wheat after canola and wheat after a pasture. Wheat after a pulse was not significantly different to wheat after a cereal. (Figure 1).  On average, wheat after canola, fallow and legume crops yielded 180,170 and 150kg/ha more, respectively, than wheat after a cereal. On average, wheat after a pasture yielded just 60kg/ha more than wheat after cereal. These rotation effects describe the average benefit that farmers receive, and are much lower than experimental break crop benefits, which are often around 600kg/ha (Seymour et al 2012)

The moderate effect of rotation on predicted wheat crop yield was important. However, the model did only explain 26% of the variation. There was considerable unexplained or residual variation across the wheatbelt and there were broad spatial trends. In the southern reaches of the low rainfall zone, yields were higher than predicted. Conversely, in the eastern wheatbelt, yields were often lower than predicted (Figure 2).

Figure 1. Crop rotation effects on wheat yields from 2017, 2018 and 2019 from 43,481 fields from the low rainfall zone of Western Australia.

Figure 2. The residual or error relating to crop rotation effects on wheat yields. High residuals imply that observed yields were higher than expected given the crop rotation and growing season rainfall. Lower residuals, that are negative, imply observed yields were lower than predicted.

The spatial pattern evident in Figure 2 can be more closely observed in Figure 3, to allow detailed investigation of the residuals. The spatial pattern could be a result of soil type, that was not accounted for in this analysis. High residuals suggest that the general explanation of break crop and water relations under estimate production. Low residuals suggest the general explanation over estimates production. Large regional differences are evident. At finer scales the graphical analysis can identify highly productive or unproductive fields, relative to neighbouring crops. Similar analyses could be conducted with other agronomic information, such as the time of sowing.

Figure 3. Detailed map illustrating where a field has a positive residual (blue colour), and is therefore producing higher yields than expected, and producing more crop than neighbouring fields. Greyed out fields did not grow wheat during the survey period.

Conclusion

The latest generation of crop models, which use satellite imagery, can now be used to evaluate age-old agronomic problems like the on-farm benefit of crop rotation. When combined with additional analysis, it is possible to determine if the farm or field in question is executing the management technique well, or could potentially improve further, relative to neighbouring fields. The difference between the size of the break crop effects in this survey are much lower than the 600kg/ha reported in the literature, which suggests that not all farmers achieve the break crop benefits observed in field trials.

Acknowledgments

This analysis was funded as part of the GRDC low rainfall zone project (CSA00056). The primary methods were developed by CSIRO through an internally funded Future Science Platform, Digiscape.

References

Donohue, R., Lawes, R., Mata G., Gobbett, D. & Ouzman, J. (2018) Towards a national, remote-sensing-based model for predicting field-scale crop yield. Field Crops Research. 227, 79-90

Seymour M., Kirkegaard J., Peoples, M., White, P.& French, R. (2012) Break-crop benefits to wheat in Western Australia – insights from over three decades of research. Crop and Pasture Science 63, 1-16

Waldner, F., Chen, Y., Lawes, R., & Hochman, Z. (2019). Needle in a haystack: Mapping rare and infrequent crops using satellite imagery and data balancing methods. Remote Sensing of Environment, 233, 111375.

Waldner, F., & Diakogiannis, F. (2020). Deep learning on edge: extracting field boundaries from satellite images with a convolutional neural network. Remote Sensing of Environment 245-111741

Contact details

Roger Lawes
CSIRO
PMB 5, Floreat, WA, 6014
Phone: 08 93336455
Email: roger.lawes@csiro.au