The paper proposes a crop identification method based on the algorithm for calculating estimates using satellites images. The classification features are the set of time series values, built by the sequence of satellite images, and geographic coordinate of field - latitude. Method was tested by using Terra/MODIS images and ground-based information. The comparison of the classification quality of proposed method with classification quality of the classifier based on Mahalanobis Distance is given.
Currently, management in agricultural sector is an area of active development and
implementation of methods that apply remotely sensed data (hereinafter - RSD) to
solve production tasks [
The literature suggests a large number of methods to solve crop recognition task by
using satellite images [
In this paper crop recognition method based on algorithm for calculating estimates is proposed. This method is suitable for usage in areas of regional scale. The method allows to take into account the geographical position of the field and to use the time series with gaps in observation days. Gaps occur due to the presence of clouds on satellite images at a certain day of observation. The ability to use data with gaps does not require an additional procedure of time series interpolation. The proposed detection method is based on an evaluation of the proximity between classified and learning objects. The quality of the proposed method is compared with the quality of the classifier based on the Mahalanobis distance. 2
All data for research were obtained for the Samara Region. The paper uses satellite images and ground data about fields for the year 2014. The description of input data and preprocessing methods is given below. 2.1
Data from satellite Terra/MODIS are used to construct time series. The raw data
recorded on the satellite are passed through ground processing and are saved in the form
of so-called products [
Ground data about fields is provided by farmers and are necessary for training the
classifier and evaluation of its quality. The following ground data is provided for each
field:
the boundary and the square of the field;
crops seeded on the field or type of field usage such as fallows or unused lands.
The boundaries of the fields were superimposed on a cover composed by satellite
images of medium resolution (20-30 m) in order to verify the homogeneity of fields.
The heterogeneity is caused by the following factors: the presence of several crops or
types of usage on the field, the use of only part of the field for sowing, sprouts
heterogeneity. All identified inhomogeneous fields were divided into the corresponding
number of homogeneous regions. The study of homogeneity was conducted using a
segmentation method [
Normalized difference vegetation index NDVI was selected for the construction of time series. Value of index is calculated according to the formula: NDVI NIR RED , NIR RED (1) channels NIR , RED ; ing to the formula (1). where NIR , RED – values of reflected radiation in the near infrared and the red regions of the spectrum, respectively.
Terra/MODIS images received during the period from March 1 to September 30 for year 2014 were used to construct the NDVI time series. An index value for each plot is calculated as follows: calculation of the values averaged by all pixels of the plot for red and near-infrared calculation of the NDVI value on the plot by averaged values NIR ,RED accordNDVI values are sorted by the date when the image was taken. Sorted NDVI values form NDVI time series of the plot. 3
To detect crops it is offered to use the classification method based on the algorithm
for calculating estimates (hereinafter – ACE). Class of recognition algorithms based
on the calculating estimates was proposed by Zhuravlev YI [
The input data for the algorithm is a set of reference objects and recognizable objects. All objects are characterized by the following features: time series values and geographical coordinate – latitude. The task is to classify a set of recognizable objects in the predefined classes. A priori information is given in a table 1. – the value of the time series of the object p in a day n ; N – the total number of observation days. Both reference objects and recognizable objects can have gaps in the values of the time series. Specifying ACE model assumes setting the following sub-paragraphs. 3.2
In this method system of support feature sets consists of a single set comprising all the features. 3.3
ed as follows: ek11k 2 , The proximity of the recognized object a and the reference object p is calculat(2) where 1 – the value that characterizes proximity of two objects by the time series values, 2 – the value that characterizes proximity of two objects by latitude, k – parameter determining the weight 1 and 2 in the final value of proximity. Values 1 and 2 are calculated by the formulas (3) and (4), respectively.
N 1 1 p,n an 2
n0 2 p a (3) (4) (5) (6) The following conventions are used in the formulas (3) and (4): an , n 0, N 1 and a – the features of object a : set of time series values and latitude, respectively. Value 1 is calculated only for the days n on which both the object p and the object a have the values of time series p,n and an , respectively.
The value of proximity function f p , a between reference object p and recognized object a is calculated as follows: f p , a 1, T ,
0, T where T – proximity threshold. 3.4
Estimate of proximity j of object a to a class j is calculated so: j
f p , a p j 3.5
1) c arg max
m0, M 1 ments in a class m .
Classification of recognized object a will be done in class c according to the decision rule. We define two variants of decision rules:
m ; 2) c arg mm0a,Mx1 qmm , where qm – the number of ele
Testing sample of 6424 plots has been formed to assess the quality of the proposed classifier. Quality evaluation has been conducted using cross-validation. Testing sample was divided five times into training and control sample in the ratio of 2: 1. 4.2
Procedure of searching the parameters k and T has been executed for each option of
the decision rule. Searching procedure allows finding such parameters k and T on
which the best classification results are achieved. Search has been made by a brute
force method with a step of 0.01 for parameter k and 0.001 for parameter T . The
best results of classification for the decision rule 1 are achieved when k 0.98 and
T 0.9945 . For decision rule 2 best results are achieved when k 0.97 and
T 0.9945 . The results of crop identification using the ACE are shown in Tables 2
and 3.
To assess the quality of the proposed classifier based on ACE the classification of the
test sample according to Mahalanobis distance [
The paper proposes a method of crop identification based on the algorithm for calculating estimates. The advantages of the proposed algorithm: the usage of time series with gaps, accounting of the geographical position of the field. As seen from the results of classification, ACE method gives the total probability of correct classification (Q = 0.72) higher than the classifier by Mahalanobis Distance (0.64). However, the value 0.72 is not satisfactory, and improvement of the quality of the classification is required. A detailed study of the testing sample showed that some of the data are unreliable: incorrect crops for some fields are given as well as the division of crops into classes was carried out nonoptimally. Therefore, the direction of future research is to study the issue of division crops into groups, so that the best classification quality is achieved at the highest possible splitting crops into classes. 6
This work was financially supported by RFBR, project № 16-37-00043_mol_a «Development of methods of using data from geoinformation systems in remote sensing data processing» and project № 16-29-09494_ofi_m «Methods of computer processing of multispectral remote sensing data for vegetation areas detection in special forensics».