direct monitoring is either not feasible or required resolution is technically not
possible, data fusion methods can help to extract necessary information. Available
data from remote sensing satellites, global positioning satellites, on-the-field nearsensing instruments, administrative registration systems, and in situ wireless sensor
networks can be processed in fusion models. In order to avoid overfitting or
underfitting problems, choosing the proper input dataset from feasible resource is
important for the training or adaptation of data fusion models.
Precision agriculture has been recognized as a promising approach for increasing
the yield efficiency. Integration of sensor-based data streams for precision agriculture and the vertical agricultural systems from the field up to the basin resource
management level is an important factor for the sustainable efficiency optimization
and the policy management. System integration costs motivate Platform as a Service
(PaaS) and Data as a Service (DaaS) solutions, including cloud computing, web
services, and wireless communication between sensors and devices. When the
necessary data that is not feasible to directly gather, proper data fusion schemes
operated on PaaS or DaaS structures enable computation of equivalent parameters
from other types of correlated existing data sets while maintaining the computational
cost efficiency.
In this chapter, we have given estimated yield efficiency mapping depending on
agrometeorological indices and remote sensing data as one of data fusion examples.
Another example is given to demonstrate how time delay neural networks can be
used to estimate root zone soil moisture. Root zone soil moisture cannot be observed
directly from radar satellites due to the penetration level of energy restriction. It
should be noted that the actual root zone soil moisture is only used in training
dataset, and neither surface soil moisture nor air humidity is used as input in this
fusion example.
Convolutional features usually improve the fusion performance of data patterns in
time or space. Wavelets are essentially based on convolutions. Deep learning with
CNN architecture also mainly depends on convolutional features. Deep
convolutional networks are a rapidly developing area in machine learning for a
wide range of applications including the data fusion.
The management of nonlinearity against overfitting is one of the key issues both
in machine learning and the regression-based data fusion methods. Increasing the
number of parameters in the big data environment requires a strategy depending on
correlation, covariance rate, and the population of the relevant training data subset.
Phenological stage-based data segmentation provides piecewise linearity in
regression-based plant growth-related estimations. R
2 value exceeds 0.8 level in
the given regional yield forecast mapping example depending on three observation
parameters gathered from remote sensing images and agrometeorological measurements. As the labeled observation data size increases on the integrated service
platforms, machine learning fusion models tend to take the place of the analytic
model-based fusion schemes. The evolution and proliferation of user application
connected platform services are expected to provide a cost-effective use of computational resources and relevant software as additional services. Multilayered service
structure brings fused spatial, temporal, and spatiotemporal data; computational
resources; GIS services; machine learning; and analytic tools together. Hence
vertical integration not only improves prediction performance, but it can also ease
136
B. Üstündağ
possible, data fusion methods can help to extract necessary information. Available
data from remote sensing satellites, global positioning satellites, on-the-field nearsensing instruments, administrative registration systems, and in situ wireless sensor
networks can be processed in fusion models. In order to avoid overfitting or
underfitting problems, choosing the proper input dataset from feasible resource is
important for the training or adaptation of data fusion models.
Precision agriculture has been recognized as a promising approach for increasing
the yield efficiency. Integration of sensor-based data streams for precision agriculture and the vertical agricultural systems from the field up to the basin resource
management level is an important factor for the sustainable efficiency optimization
and the policy management. System integration costs motivate Platform as a Service
(PaaS) and Data as a Service (DaaS) solutions, including cloud computing, web
services, and wireless communication between sensors and devices. When the
necessary data that is not feasible to directly gather, proper data fusion schemes
operated on PaaS or DaaS structures enable computation of equivalent parameters
from other types of correlated existing data sets while maintaining the computational
cost efficiency.
In this chapter, we have given estimated yield efficiency mapping depending on
agrometeorological indices and remote sensing data as one of data fusion examples.
Another example is given to demonstrate how time delay neural networks can be
used to estimate root zone soil moisture. Root zone soil moisture cannot be observed
directly from radar satellites due to the penetration level of energy restriction. It
should be noted that the actual root zone soil moisture is only used in training
dataset, and neither surface soil moisture nor air humidity is used as input in this
fusion example.
Convolutional features usually improve the fusion performance of data patterns in
time or space. Wavelets are essentially based on convolutions. Deep learning with
CNN architecture also mainly depends on convolutional features. Deep
convolutional networks are a rapidly developing area in machine learning for a
wide range of applications including the data fusion.
The management of nonlinearity against overfitting is one of the key issues both
in machine learning and the regression-based data fusion methods. Increasing the
number of parameters in the big data environment requires a strategy depending on
correlation, covariance rate, and the population of the relevant training data subset.
Phenological stage-based data segmentation provides piecewise linearity in
regression-based plant growth-related estimations. R
2 value exceeds 0.8 level in
the given regional yield forecast mapping example depending on three observation
parameters gathered from remote sensing images and agrometeorological measurements. As the labeled observation data size increases on the integrated service
platforms, machine learning fusion models tend to take the place of the analytic
model-based fusion schemes. The evolution and proliferation of user application
connected platform services are expected to provide a cost-effective use of computational resources and relevant software as additional services. Multilayered service
structure brings fused spatial, temporal, and spatiotemporal data; computational
resources; GIS services; machine learning; and analytic tools together. Hence
vertical integration not only improves prediction performance, but it can also ease
136
B. Üstündağ