The study area (Figure 1) is the Western Cropping Zone (WCZ) within the State of Queensland, Australia, which is based on a strategic cropping land map by the Queensland Department of Natural Resources and Mines [39]. The WCZ major industries are mining (particularly for coal and coal seam gas) and other agricultural sectors (i.e., cropping and livestock grazing). QLUMP maps 9.7% of the total area of the western cropping region as “cropping” land, with 85.4% mapped as grazing and 0.4% as mining.

The area is summer-rainfall dominated (350–500 mm) originating from prevailing south-easterly winds, with average annual precipitation between 500 and 700 mm and a high between-season variability [40]. Cropping land use is mainly determined by the presence of highly fertile clay soils (classified mainly as Vertosols, according to the authors of [41]) and well-drained, friable clay-loam soils high in iron (ferrosols), which occur in patches [40].

Queensland’s gross value of broadacre cropping commodities for 2013–2014 were 2.7 billion dollars [42]. Over half of this is produced in the WCZ (the remainder being the coastal 1.2 billon dollar sugarcane industry). The dominant summer crops are sorghum, maize and cotton or (in the case of late summer rain) mungbeans. The dominant winter crops are wheat, barley, and chickpea [43]. Due to a strong north-south climatic gradient, the WCZ can be subdivided into “NW” (for north-western) and “SW” (south-western) regions. This gradient and division is mainly driven by increased evaporation due to latitudinal radiation increase and lower winter rainfall in the NW [40].

Some large areas of the WCZ’s most productive cropping land coincide with areas of economically viable coal. This conflicting land use has implications for food security and the long-term economic potential from agriculture. In 2014, the State Government of Queensland introduced legislation to help manage the competing land uses in areas of regional interest, including the WCZ. The Regional Planning Interests Act (RPI) 2014 aims to protect high-quality agricultural areas and areas considered to be strategic cropping land [44]. To help identify these areas, current and historical spatial information is required, particularly information and data which can identify how often, and where, cropping activity has been undertaken.

The WCZ primarily has a winter and a summer growing season [40]. We therefore aim to identify actively growing crops in these key growing seasons. For the purposes of this study, we define the summer growing season as November to May (inclusive), and the winter growing season as June to October (inclusive). When attributing a year to a summer-growing crop the most recent was used, e.g., “summer 2013” represents the period November 2012 to May 2013.

All available Landsat 5 TM, Landsat 7 ETM+ and Landsat 8 OLI data across 27 footprints with cloud cover of less than 60% were used (Figure 2). The archived images over Australia appear only with higher image frequency since late 1987 [45]. This criterion was chosen to reduce the potential for noise in the data due to imperfect cloud/cloud shadow masking and water vapour variations.

All images were cloud-masked with a regionally adapted version of Fmask [46,47,48] then NBAR-corrected [46]; this latter step converts the units of the imagery to bottom-of-atmosphere reflectance, scaled to the interval (0, 1). The high internal geometric consistency of the USGS-based Landsat imagery allows further data processing and analysis without further geometric adjustment [28].

All archived images were used to generate a synthetic spectral image for each season (s) and year (y). The synthetic images are denoted herein L ( s , y ) . The images were generated per pixel and per spectral band (excluding the thermal band), based on a statistical model [49] (Figure 3). Seasonal mid-points, t₀, of cropping activity were defined as 14 February for the summer growing season and 15 September for the winter growing season. These mid-points represent the times when generally the most plant vigour is expected in the study area [43].

Denote the time series of reflectance for a particular Landsat pixel and band within L ( s , y ) i , j with i representing the number of images within a season (i = 1, …, n); and j a single pixel. When an observation coincided with t₀ nothing more needed to be done; the reflectance value entered the synthetic image directly. Otherwise, when n = 1 a nearest-neighbour interpolation was applied to establish the reflectance value for t₀; when n > 1 we calculated, in statistical parlance, the “location” and “scale” of L ( s , y ) . For 2 ≤ n ≤ 4 location and scale were taken as the conventional mean and standard deviation, respectively; for n > 4 we invoked robust statistics, where location and scale became the median and median absolute deviation of the reflectance z at t₀. The median absolute deviation is given by the authors of [50] and gives a robust estimation of one standard deviation: (1) scale ( L ( s , y ) ) = 1.4826 { median ( | L ( s , y ) i , j − median ( L ( s , y ) ) | ) }

Any observation outside the interval: (2) location ( L ( s , y ) ) ± 2 { scale ( L ( s , y ) ) } was considered an outlier and adjusted to the interval (Figure 3). This procedure is known as Winsorising [49]. A linear regression was then fitted to the nearest four observations to t₀, with z as the response variable (reflectance), and time as the single explanatory variable. Note that Winsorising does not remove the effect of an outlier, it merely reduces its influence on subsequent analyses.

The most recent QLUMP mapping was used to mask areas from each synthetic image that were associated with land uses that were not of interest for this study (e.g., conservation areas, forestry, and intensive uses). Travelling stock routes, which are areas reserved along roadways for the movement of livestock, were also masked from each image. Water bodies were masked for each synthetic image by calculating the water index of the work in Reference [51] for each image in L ( s , y ) . If the minimum of the water index was > t_r (the recommended threshold was t_r = 68 units) it was assumed that the pixel was not covered by water during the growing season and hence left unmasked; otherwise, a one-sided Wilcoxon signed-rank test was used to determine if the median water index over the growing season was < t_r. If the null hypothesis was rejected at p < 0.05 the pixel was considered inundated for a substantial part of the growing season and masked from all further analyses.

To reflect the fact that land is managed in discrete parcels, and to reduce data loads, the (masked) synthetic images were segmented into homogeneous areas using the open-source software package RSGISLib with Python bindings [52,53]. Each synthetic image was segmented independently to account for change between seasons. The segmentation combines a k-means clustering algorithm with a clumping step to find the connected regions of the image [53]. The number of cluster centres and the minimum segment size are required input parameters. A minimum segment size of 50 pixels (4.5 ha) with 60 k-means nodes per Landsat image were used (recommended settings; [52]). Sixty nodes appeared to work best on a full Landsat image, while the minimum segment size of 50 pixels was empirically determined as a good compromise between avoiding an under splitting and too much over splitting. The segmented image was stored in the GDAL-supported KEA image format [54] where the image segments are linked with a raster attribute table (RAT) and used as the basis of all further analysis. This step is a method of data reduction by moving away from a pixel-based analysis to spatial medians per segment for all classification variables. A final step used a 1-s digital elevation model [55] to mask out segments with a slope of greater than 10% as cropping is generally not undertaken in steep terrain within the study region.

A range of variables were calculated in order to capture reflectance signatures and temporal characteristics, which were used to classify the image segments (Table 1). Firstly, the mid-season values for a number of vegetation indices and band ratios were calculated for the pixel-based mid-season synthetic image. Similarly, pixel-based NDVI values were calculated from the mid-season synthetic image, as well as descriptions of several temporal metrics (Table 1). The RATs were populated with the spatial median per segment for each variable listed in Table 1.

A final variable was created based on the land use data from QLUMP [24] to help constrain the classification to known areas of cropping. All areas classified as “cropping” were extracted from the QLUMP mapping data to create a mask of cropping areas. It is important to note that, due to the limited temporal information that is used for mapping cropping areas in the QLUMP, these areas classified as cropping may not be a comprehensive spatial representation of all areas of active cropping over time. Hence, while being a potentially useful explanatory variable for classification purposes, the land use mapping was not used as an explicit representation of all cropping in the study area. The metrics used here are specifically targeted to distinguish the Crop/No-Crop classes and differ from, e.g., the authors of [8,30] who aimed for a more general land use classification.

Field-based training data described in Reference [43] was used in this study. These data, collected from various campaigns between September 1999 and February 2012, include observations about presence and absence of crop (e.g., wheat, maize, sorghum, etc.) at the time of observation. For the purpose of this study, all records were recoded as either Crop or No-Crop. The field data were intersected spatially and temporally with the image segments from each synthetic mid-season image. When more than one observation fell within the same image segment only the first was retained and only if the observations belonged to the same class (Crop or No-Crop); if they had different classes none of the observations were used.

An additional set of records were generated through desktop interpretation of the seasonal synthetic images. The sites were sampled at random over space, growing seasons, and years. The intention of these records was to provide an unbiased assessment of cropping activity that could serve as validation data. From a desktop, an expert would “visit” a sampled location in space-time and use L ( s , y ) data (and any other higher-resolution imagery if available) to assess whether the on-ground activity was either Crop or No-Crop. Occasionally the sampled spatial coordinate fell on a field boundary or road; in this case, the coordinate was manually moved inside the nearest field (but never more than a distance greater than 3 pixels).

These records form a set of explanatory variables and a response variable amenable to statistical modelling. A breakdown of the numbers of training and validation data is shown in Table 2. In NW 51% and 52% for summer and winter, respectively, were derived from the collated field data. For SW the field data accounted for 49% and 46% of the total data set for summer and winter, respectively.

The number of validation data points for the class Crop is relatively small (Table 2). The random sampling approach (with 9.7% of the study area being mapped as “cropping” by the QLUMP) does imply that about 10% of the validation data should be Crop if all areas are actively cropped at all times. The class Crop is, however, in the validation dataset represented with an even smaller percentage, resulting in an unbalanced dataset. Table 2 reflects this distribution and implies that a validation based on these data is, due to the small number of points for the class Crop, potentially misleading. Due to the intensive field sampling and farm surveys, more data on the actively cropped fields is available in the training data. Therefore, for the purpose of training and validation, the original validation data were merged with the original training data. A 75% random subsample of the merged data was chosen for the model training and the remaining 25% subsample quarantined for validation. Thus, a total 1387 points were used for validation in summer and 1359 in winter across the regions.

The generation of a categorical classification generally has the underlying assumption that the features in question can be systematically separated into distinct classes by their numerical properties. The classes in a supervised classification are specified a priori, and associated with a set of training data for each class. A wide range of algorithms exists to choose from, each with their specific advantages and disadvantages. The authors of [59] tested 179 different classification models arising from 17 classification families on 121 different datasets, to answer the question if we always need to apply multiple classifiers and choose the best performing algorithm. They came to the conclusion that Random Forest (RF) [60] is most likely performing best, but with no significant improvement over the second-best performing algorithm, Support Vector Machines (SVM) [61]. In this approach we compared four classification algorithms with the same training data: (i) SVM; (ii) multinominal logistic regression; (iii) the decision-tree classifier C5.0 ([62]; and (iv) RF; all within the R software. Table 1 describes the explanatory variables spanning the classification feature space. A total of 23 variables were generated for each region and season.

In a (non-linear) SVM classification the feature space is transformed to a higher dimension using what is called a “kernel trick”, where a hyperplane is fitted to separate the different classes [61,63], which may not be feasible in the lower dimensional space.

For each of the classifiers used, the probabilities of the class membership were stored per image segment. Each segment was labelled as Crop if the predicted probability for that segment was greater than 0.5, due to the binary nature of the classification.

The classification is performed ensemble for all years per season and region, to account for inter-annual variability in the cropping and for years without field data.

Figure 4a,c provide examples of the mid-season synthetic images for a subset of the SW region for the winter growing season and the summer growing season, respectively. Figure 4b,d show the respective single-date images closest to the mid-season date (in this case both Landsat 7). Both images contain clouds with different degrees of image contamination with the Fmask based cloud and cloud shadow polygons overlaid. The robust seasonal synthetic image algorithm appears to work well; even if at times artefacts can be found, e.g., if scan-line correction (SLC) is off in Landsat 7 ETM+ or if the temporal gap between images is large (i.e., due to changes in vegetative phenology).

The temporal behaviour of a crop in relation to the synthetic image is graphically illustrated in Figure 5 for a centre pivot circle (yellow outline in Figure 4a), which is fallow in summer 2014 and irrigated in the following winter 2014. The green data point marks the predicted value at the mid-season date, while the red data points represent Landsat 8 OLI and the black data points Landsat 7 ETM+ spatially averaged (over the extent of the circle) data values.

Figure 5 demonstrates that the prediction point for summer is well aligned with the Landsat observations, while prediction point for winter 2014 has some deviation, particularly in Bands 1 to 3, due to the effects of clouds and SLC-off. Generally though, the predicted green points are in line with the temporal behaviour of the curve. This suggests that Bands 1 to 3 are likely to have been affected by other factors including variations in atmospheric constituents, such as water vapour in decreasing magnitude [46,47]. The synthetic image generation based on the Winsorising approach appears to have filtered the effects of contamination, and retained the overall trend.

The quality of the synthetic image generation was evaluated in an image subset for scene 90/79. The synthetic image was generated without the image observation on 15 September 2015 (Figure 6a). The synthetic image (Figure 6b) resembles the real image fairly well as the difference map (Figure 6c) indicates. The major differences visible are a crop circle, where the band 4 reflectance was overestimated (lower left) and some close by fields which were underestimated. Both of these cases are due to the phenological changes in the latest image selected in the prediction (9 October). The majority of the differences in band 4 is between ± 0.05, which is similar to the synthetic image generation by the authors of [35]. A regression between the observation and synthetic image revealed a residual standard error of 0.021 with an adjusted r² of 0.95 (p < 0.00001).

Despite the (minor) differences in greenness was the synthetic image considered adequate for GEOBIA.

The spatial consistency of the cloud-free synthetic image enables the application of an image segmentation approach. Each synthetic image over time is segmented independently, as field sizes and shapes may vary. Figure 7 illustrates the result of GEOBIA over a complex and heterogeneous area within the same image subset as used in Figure 3. The amount of strip cropping seen here is an extreme example of image heterogeneity within the study region.

The majority of the landscape features and field outlines are captured by GEOBIA. Some larger fields are represented boundary sharp as individual segments. The segmentation parameters appear to not separate the strips at the highest level of detail, e.g., if the strip cropping is very narrow (i.e., 1–3 pixels wide). Within-field variation can cause the algorithm to split one field or pivot circle if the heterogeneity is high [33]. This over splitting issue does not pose a problem for the overall approach as the spatially aggregated values are classified as Crop or No-Crop only. The median number of segments per Landsat path/row in NW was 74,514 and 79,861 in SW, with a median area of 873 ha and 879 ha, respectively.

The image segments for the cropped field may not always be sharp on the field boundaries so that some discrepancies with the true field outlines are apparent. This may especially when viewing the classifications over time result in some fuzziness at these borders.

Figure 8 represents an example of the image classification for several seasons, based on the (RF + landuse) classification. The full time series of the subset is shown in the Appendix as Figure A1. The same image stretch has been applied to the mid-season synthetic images, while the probabilities for active crop are colour-coded according to the legend in Figure 8 (yellow to green). The comparison of the mid-season synthetic image is effective if the crops are greening-up during this particular time, but if a crop is planted late, e.g., due to a late summer rain, the crop might not appear green in the synthetic image. This appears to be the case for the majority of fields, but for an accurate assessment all of the seasonal imagery have to be assessed as the general peak vegetation date does not necessarily coincide with the real maximum of vegetation greenness.

Table 3 shows the accuracy statistics for each of the five classification approaches and for each region and growing season. The same training and validation data were used for each classifier. It is important to note that the classifiers were trained and validated over all years for which data were available.

In some instances gave SVM and Logit the superior classification results, but overall the RF classifiers were always amongst the better performing. The inclusion of the prior information derived from land use mapping improved the classification accuracy considerably for summer, especially the user’s accuracy of the class Crop. To visualise the variable importance the MeanDecreaseGini metric [60,64] is displayed (Figure 9). This is a measure of variable importance based on the Gini impurity index used for the calculation of splits during RF training. Every time a node split is made in the RF based on an explanatory variable, the Gini impurity criterion for the two descendent nodes is less than the parent node. Adding up the Gini decreases for each explanatory variable over all trees in the forest gives an importance measure that is often very consistent with the permutation importance measure [60,64]. Figure 9 plots the 10 most important variables in the four cases (NW and SW in summer and winter).

In the NW subzone, for summer and winter, there are 6 temporal metrics within the 10 most important variables; in the SW subzone 5 out of the 10 most important were temporal metrics. It is also revealed that, in summer, prior knowledge of land use is an important aspect of the classification.

The (RF + landuse) classification was applied to all seasons and years. Figure 10 shows an example of the crop probabilities predicted for the WCZ for the summer and winter growing season of 2014.

The classified surfaces are available for all season and years (1988–2015) for download via the Terrestrial Ecosystem Research Network website ([67,68]).

Summarising the active crop history over a defined period of time can give an insight of the spatial distribution of the crop activities and intensities of use over last 27 years. This is visualised in an image subset in Figure 11, where the intensive use in the summer irrigation near a river stands out in the long-term cropping count (Figure 11f).

Figure 11 also indicates the different land use interests with the open-cut coal mine Acland mining lease of stage 1 in 2005 outlined in blue and subsequent extension leases including a rail line. The stage 3 extension was approved (with conditions) in 2012 and is currently under review by the Queensland Department of State Development Infrastructure and Planning [69] which entails an environmental impact assessment. The use of the historic imagery and in particular the active crop frequency layers will be investigated for updating the existing land use data.

A field that was not part of the most recent land use layer as outlined in Figure 12 was captured as cropped multiple times in the 20-year period of 1995–2014.

The image segments for the cropped field may not always be sharp on the field boundaries so that some discrepancies with the true field outlines are apparent. This may especially be the case when viewing the classifications over time, resulting in some fuzziness at the borders.

The synthetic image generation step is vital for the subsequent analysis as the image segments and some of the indices used in the classification are based on this. For the majority of seasons there are 10 or more Landsat images with less than 60% cloud cover. The threshold was applied to reduce the risk of image artefacts from imperfect cloud and cloud-shadow masking, thin undetected clouds and water vapour variations which can influence an individual pixel’s reflectance values. Only a few footprints in a specific growing season have fewer than five images available. Too few images per season may lead to artefacts in the synthetic image generation and unrealistic temporal metrics and thus reduce the overall accuracy. This could only be overcome in some regions by carefully selecting and including additional scenes with higher cloud cover, or if further Landsat imagery would be available from yet unknown Landsat archives.

The GEOBIA used here is well-suited to both the broadacre cropping areas and the smaller, more heterogeneous fields (Figure 7). Over-splitting of the segments can occur due to within-field spatial heterogeneity, or in cases where the base image displays stripes originating from ETM+ SLC-off data. The over-splitting, however does not pose a problem to the classification as these polygons are attributed and classified separately. Under-splitting may occur, if the field sizes due to strip-cropping become small (e.g., a few pixels in width) or, e.g., if field boundaries are diffuse [33]. A more detailed segmentation could help to separate these strips further. This could include a regionally adapted parameter setting (e.g., for the minimum segment size), but would most likely require higher spatial resolution imagery. This may become important for a crop type classification, but in our case for a binary Crop/No-Crop classification this is not as relevant, so one setting with a large number of clusters per Landsat footprint was used in the segmentation. The authors of [32] performed an “individual crop field” accuracy assessment with reference polygons while using a five-year composited image product to gain more defined image boundaries on repeatedly cropped fields. Accurate field delineation is not a primary focus of this study, but outputs of this classification are envisaged to feed into the QLUMP system to aid the updating of land use and potentially field boundaries.

The approach applied here is designed for large-area mapping and our results indicate that mapping at field-scale works with acceptable accuracy. In the spatial aggregation step, spatial median values have been calculated, making the approach more robust against spatial outliers such as pixels affected by imperfect cloud or cloud-shadow masking. One of the benefits of the GEOBIA is that there is no removal of “salt and pepper” effects required [33,37].

NDVI was used as the primary vegetation index for the temporal metrics. The Enhanced Vegetation Index (EVI) is commonly used for vegetation studies [70] to reduce the effect of soil. As we are not comparing absolute values of a vegetation index across large spatial extents, rather the temporal signal in a relative manner over time, NDVI is just as useful, as the soil is not changing drastically over time for each location (three decades). A similar argument applies for a potential cross-calibration of the Landsat 5/7 with Landsat 8 NDVI, which was discarded in this case: the authors of [46] investigated the difference in NDVI and found an up-to-5% discrepancy between Landsat 7 and 8. When applying a correction factor the difference was reduced to 1% across all Australian natural landscapes. This suggests that for NDVI estimates, the difference between the two instruments is small. The difference may yet be smaller in the cropping systems with different biophysical properties so that an improvement is not guaranteed at this stage, but should be investigated in the future.

A vegetation cover description with an entire different approach, such as spectral mixture analysis [71] may add additional information and will be investigated further. The temporal metrics are, however, among the most important variables in the classification.

The comparison of classification algorithms revealed that the SVM classifier and the RF classifier perform similarly. This is supported in the findings of Reference [59] and, following their recommendation and the results obtained from the classification, the RF classifier was used. While a particular classifier may perform slightly better (i.e., SVM) for a season or year, one classifier was chosen for all cases for the purpose of processing speed, comparability and consistency, given the large areas and longer time series over which the approach was to be applied.

We found that, in the absence of land use information, it was difficult to predict Crop with adequate accuracy in summer. This is because annual pastures tend to grow with summer rainfall events, resulting in increased greenness, and hence classification confusion. Similarly, in years with unusually high rainfall (e.g., the winter growing season of 2007), an overestimation of cropping activity occurs as pastures can show a rapid vegetative response. Actively growing crops can therefore be confused with pastures or improved pastures, particularly where the pasture growth is further promoted by irrigation (e.g., for hay production). This is reflected in the improvement in the (RF + landuse) model in the users accuracy for the class Crop in summer for both regions (about 14% in NW and 3% in SW). The class probabilities for the class Crop were mapped and colour-coded. A probability threshold of 0.5 was applied globally as a demarcation in the binary classification. In some areas a local adjustment of the probabilities might result in a better crop differentiation. The inclusion of land use as an explanatory variable comes at no real cost to the model and, with the exception of the NW winter (where the producers accuracy is reduced by 2%), the overall accuracies are improved, especially in the summer season (Table 3) i.e. user’s acc. in NW summer. In the case of a large time difference between the land use mapping and the crop-mapping date, a false weighting may result in mis-classification. Incorporating historic land use maps, where they exist, may reduce this effect, although historic maps may differ in scale and accuracy compared with more recent mapping. It could be argued that an accuracy improvement of about 1% in NW winter or in SW winter does not warrant the (RF + landuse) model over RF, however for consistency, the same model was used for each region and season. While the land use mapping was included here as a variable, the classification results suggest that in the absence of a land use map, the RF approach would still achieve reasonable results.

The classification accuracies based on the Kappa statistic are between 0.88 and 0.97 suggesting that a good separation between the grazing land and active cropping can be achieved. More validation data are desirable, especially for the class Crop, where the data set is unbalanced due to the small proportion of the cropping in the study area (9.7% according to the land use map). The statistics reveal that in the NW sub-zone the producer’s accuracy for the class Crop in the summer growing season is the lowest at 0.88, while for NW and SW summer the user’s accuracies for Crop are the lowest with 0.90 and 0.91 (Table 3).

Based on our mapping of cropping activity, further developments are planned. This includes testing a crop-type classification, and then extending that to yield mapping and forecasting through the incorporation of climate information and other sources of Earth-observation data, such as Sentinel-1 and 2. Gaps in the image archive can pose a problem for the procedure, particularly when the 16-day gap between Landsat observations coincides with a period of dense cloud cover. Regularly spaced observations would allow for a phenology-based approach that also has the potential for crop-type differentiation [43], but might require an image-fusion step to be applicable at field-scale—for example Landsat with MODIS data [72]; [73] or Rapideye and MODIS [74]. The authors of [75] demonstrated the use of multispectral Landsat information combined with temporal MODIS information for crop mapping. The authors of [76] used a MODIS- and Landsat-fused time series to investigate tillage management practices in agricultural systems. A potential drawback can be the loss in spectral “sharpness” [76,77] in the fused data and the large data volume. The maps produced are available online an can be used as independent data to validate coarse-resolution crop mapping approaches at a continental or global scale [16].

The advent of the European Space Agency’s Sentinel-2 imagery is expected to increase the classification accuracy, as more observations per season are expected combined with a finer spatial resolution. The classification procedure is based on simple vegetation indices and band ratios so that data from different sensors can easily be integrated and new seasonal crop maps can be generated. Further, the integration of Landsat Multi Spectral Scanner data would extend the crop mapping approach back to 1972, but larger temporal gaps in the data may require a different approach [78].

The data generated and the framework as such could be considered for crop mapping approaches within GEOGLAM (Group on Earth Observations Global Agricultural Monitoring Initiative).