Empowered by geo-locating and sensor-based technologies, precision agriculture brings a data-intensive paradigm into farming. In this spirit, we investigate the role of outlier detection and visualization in decision-making for precision agriculture. We discuss two analysis tasks for visually monitoring fields that exhibit problematic crop growth compared to their neighbors, and for visualizing problematic areas inside a field. As a proof of concept, we analyze satellite imagery to case-study our tasks in the context of the Future Cropping project.
The proliferation of the Global Positioning System (GPS) and of
sensor-based technologies have given rise to a diferent paradigm
of agriculture. Precision agriculture as this paradigm is known,
strikes to improve economic returns and reduce the
environmental impact [
Anomaly or outlier detection has been widely studied in data
mining to support decision-making; the goal is to identify objects
that significantly deviate from the expected pattern in a dataset
[
In precision agriculture, outlier detection is mainly studied for
pre-processing and cleaning of the collected data [
In an attempt to fill this gap, we propose the tasks of inter-field
(i.e., between fields) and intra-field (i.e., within a field)
analysis. In the former, our goal is to identify fields, which exhibit a
problematic behavior, e.g., unexpectedly low collected crop mass,
compared to their neighboring fields. 1 The key challenges for
this task are twofold; (i) how to define field neighborhood and (ii)
how to quantify the importance/influence of each neighbor.
Different from traditional spatial outlier detection that relies solely
on arbitrary polygon distances, we consider domain-specific data
properties such as soil/crop type and the variation of crop-growth.
On the other hand, intra-field analysis is triggered in an efort to
interpret the results of inter-field analysis. The goal is to identify
and visualize the areas inside a particular problematic field that
significantly deviate from the average or normal behavior of the
ifeld. Practically, these are the areas where the farmer may need
to take action [
Recent studies [
Besides spatial geometries (e.g., polygons and/or point data), our outlier detection methodology uses for every field (or parts of it) a set of numerical attributes that store measurements such as the collected crop mass, and a set of categorical attributes which store properties of the field such as its soil and crop type. Our analysis focuses on a particular numerical attribute of interest denoted by α , which monitors the crop growth on a field; Table 1 summarizes the notation used throughout the rest of this paper.
We model inter-field analysis as a spatial outlier detection task
[
Conceptually, the process of inter-field analysis involves three steps. First, the neighborhood of each field f is defined on the basis of the contextual attributes. Under this, selecting the proper distance measure dist between two fields is critical; dist takes into account the spatial proximity of the fields and potentially the values of other contextual attributes, e.g., the soil type. Let Nf be the set of neighbors for a field f . The second step is to predict the value for the attribute of interest α on f by aggregating the values on f ’s neighbor fields in Nf ; typically, the mean value is used for this purpose. 2 In our study, we calculate f .αb as the weighted mean value, which allows us to weight the importance and the contribution of each neighbor, considering both the spatial distance to field f and other factors as discussed in Section 3.2. Formally, we have: f .αb = Σi|N=1f | {wi · ni .α }
Σi|N=1f |wi
where wi denotes the weight of neighbor field ni . The third step
is to report the outlying fields. The value of | f .αb− f .α | quantifies
the deviation of α ’s predicted value from the actual value on a
ifeld f . In practice, there exist two approaches for determining
outliers. We can either rank the fields in decreasing order of their
| f .αb− f .α | deviation and output the top k, or return all fields with
an extreme deviation value. Probabilistic and statistical models
can be used to determine such extreme deviation values [
The above procedure allows us to identify fields whose predicted value on attribute of interest α significantly deviates from the actual value. In fact, we distinguish between two types of outliers based on the sign of the f .αb− f .α deviation. If f .αb− f .α > 0 2As an alternative, median can be used to smooth the impact of extreme α values. (1) then field f behaves worse than expected based on its neighborhood; we call f a negative outlier in this case. Otherwise, f is a positive outlier that exceeds our expectations. Both cases are important for decision-making in precision agriculture as negative outliers may require immediate action by the farmers and positive outliers may provide important feedback to improve crop growth in other fields.
Conceptually, the process of intra-field analysis involves three steps. First, the extent of each field is split into a number of partitions; each partition represents an area of interest. In practice, such a partitioning may for instance be provided by the farmer or generated by a partitioning process. For instance, in our case study of Section 3, we used the image resolution to partition our satellite data. Similar to the entire field, we assume that each partition is associated with the same attribute of interest α . 3 We next define |pi .α − f .α | as the deviation of α ’s value on a particular partition pi , from the value on the entire field f . Last, we can employ the two approaches discussed in the previous subsection for the inter-field analysis to report the outlying areas of a field, i.e., either by ranking the partitions by their |pi .α − f .α | deviation value, or by reporting the partitions with an extreme deviation value.
We case-studied our outlier detection tasks for the inter- and the intra-field analysis in the context of the Future Cropping project, which collects and analyzes agricultural data in Denmark. 4 Our study involves two types of data; (i) the geometries (i.e., the polygons) of all 590,490 fields in Denmark and their crop type, and (ii) a 6.3GB vegetation map derived from ESA Sentinel-2 satellite imagery captured on May 8, 2016. 5 We conducted our case-study on QGIS. 6
In the past, a number of vegetation indices [
N DV I = N I R − V I R (2)
N I R + V I R where N I R and V I R are the near-infrared and the visible red bands, respectively.
For our analysis, we use N DV I as the attribute of interest α . Specifically, we first calculate the index value on every pixel of the satellite images using Formula 2. Then, for each field f , we compute mean µ f and standard devation σf of its N DV I values based on the overlapping pixels, and set f .α = µ f . Figure 1b 3Essentially, the value of attribute α on both the entire field and a partition is computed by aggregation. For instance, harvesters store position-aware crop data for the collected mass on specific geographic locations inside a field; we can aggregate these values to compute the overall collected mass for the entire field or for particular areas. 4https://futurecropping.dk 5https://kortdata.fvm.dk/download/Index?page=Markblokke_Marker, https://scihub.copernicus.eu 6https://qgis.org (a) (b) We now present our case-study for the inter-field analysis and exemplify its results. We apply Formula 1 on each field f in Denmark to predict its N DV I value, i.e., f .αb. For performance, previous works in outlier detection typically define neighborhood Nf as the k closest fields. However, these neighbors (or part of them) could be in fact very distant to field f , which will potentially afect the quality of the results. Instead, we narrow the extent of a neighborhood; hence, Nf contains all fields of the same crop type as f within a radius of 10km, as suggested by our agricultural project partners. Formally, we define the distance of f to a neighbor field ni as: (
if fields f , ni have same crop type ED(f , ni ) MAX _DIST otherwise (3) where ED(f , ni ) is the Euclidean distance of fields f , ni polygons and MAX _DIST is the maximum 10km distance.
To weight the importance of a neighbor ni ∈ Nf , we take into
account two factors; (1) its spatial distance dist (f , ni ) to f and
(2) the variation of crop growth inside the field. 7 Essentially, we
prioritize neighbors located close to field f , with crops evenly
7The importance of low-variation fields has been previously studied e.g., in [
dist (f , ni ) wi = 1 − MAX _DIST · [1 − σi ] (4)
With Formulas 1, 3 and 4, we compute the f .αb− f .α deviation of the predicted value from the actual f .α N DV I value, on every ifeld f . As discussed in Section 2.1, one option for our inter-field analysis, would be to rank all fields by their deviation and recommend the ones with the k highest values as potential negative outliers and those with the k lowest deviation values as positive outliers. However, such recommendations have little practical merit for the end-user; for instance, even the highest f .αb − f .α value may in fact be too low to mark field f as problematic. Instead, we focus on identifying the fields that exhibit the most extreme deviation values.
Figure 1a reports on the distribution of the f .αb− f .α values for
a particular type of crop called winter rapeseed. The mean value
and the standard deviation for this distribution are µ d = 0.0013
and σd = 0.1645, respectively. We observe that over 99.5% of
the fields have a f .αb − f .α value inside [µ d −3·σd , µ d +3·σd ]. A
similar conclusion was drawn for the rest of the crop types in
our case-study. Based on this observation, we follow a typical
statistical approach for identifying extreme f .αb− f .α values [
Figure 1b visually exemplifies our inter-field analysis for the winter rapeseed crop; the negative outlier field f is highlighted by a red border while a subset of its neighborhood is given in black. Further, Table 2 lists the characteristics of the 10 most important neighbors of f , sorted by their weight wi (Formula 4); notice how the closest field to f , n7, is not its most important neighbor as n7 exhibits a higher variation of crop growth compared to fields n1 to n6. For field f , we have its actual N DV I value f .α = 0.176 while using Formula 1, we predict that f .αb = 0.671. The extreme deviation of the predicted from the actual N DV I value is captured in Figure 1b where the majority of the shown neighbors are 8Note that µ d , σd are statistics computed for the particular crop type of field f . significantly greener than field f , indicating that f ’s actual N DV I value should have been higher than 0.176. 3.3
We briefly discuss our case-study for the intra-field analysis. As mentioned in Section 2.2, the first step is to partition a field. For this purpose, we use the pixel grid of the satellite imagery in order to best exploit the 10m × 10m resolution. Figure 2a illustrates the partitioning of a winter rapeseed crop field; each partition (grid cell) pi corresponds to a pixel of the N DV I map and its pi .α is set to the N DV I value of that pixel. For every partition pi , we then compute the |pi .α − f .α | deviation of its N DV I from the value of the entire field. Last, we employ the approach of Section 3.2 to find outliers. Figure 2b visualizes the results of our analysis; the |pi .α − f .α | value of the partitions drawn in the figure deviates from mean µ d , at least 3 times the standard deviation σd . Observe how our analysis is able to report the partitions that cover the light colored problematic areas of the ifeld. Additional problematic areas can be identified by reporting partitions with a |pi .α − f .α | value deviating 2 times, as shown in Figure 2c. 4
We studied the role of outlier detection in decision-making for precision agriculture and how such tasks can benefit from processing satellite imagery. For this purpose, we discussed the tasks of inter- and intra-field analysis, addressing the special challenges raised by the agricultural domain. As a proof of concept, we conducted a case-study supported by ESA satellite images within the Future Cropping project.
In the future, we plan to extend our study in multiple directions; (i) incorporate more types of data, e.g., collected from harvesters or rain distribution data, (ii) employ visualization tools to collect and evaluate farmers’ feedback, (iii) enhance intra-field analysis for recommending explicit actions to farmers, (iv) further analyze satellite imagery to monitor the outlying fields or problematic areas through the course of time and (v) investigate techniques from Machine Learning for predictive analytics.
This work was funded by Innovation Fund Denmark as part of the Future Cropping project ( J. nr. 5107-00002B).