reduces the spatial dimension, while the depth does not change. Hence, the amount
of parameters is reduced, and this reduces computational cost. It also controls
overfitting.
Deep learning is not only used for the classification of one data type as cropland
cover (Kussul et al. 2017); it also enables fusion of the different types of data layers
either by applying a different kind of decomposed datasets into common fully
connected network as similar to Fig. 7.21 or they can also be merged before the
convolutional decomposition layer.
Residual neural network (ResNets) (Kaiming et al. 2015) is a deep learning
structure that improves the classification performance by utilizing skip connections
or shortcuts to jump over some layers. DenseNets were proposed in 2016, and they
use several parallel skips as an improvement of ResNets. Deep learning provides an
efficient solution in the registration of relevant data input set features that includes
the projection and rotational differences. For example scale-invariant feature transformation (SIFT) algorithm was used to solve especially this feature matching
problem in the past years. Object features are efficiently matched in deep learning
structures without the necessity of additional preprocessing methods (Sachdeva et al.
2017). This property makes deep learning as an efficient candidate, especially for
spatial and spatiotemporal data fusion processes. A widespread implementation
method of the deep learning services is using dedicated frameworks. Tensorflow,
Keras, Pytorch, Caffee, Theano, Apache MXNET, and Microsoft CNTK are widespread deep learning frameworks by the year 2018. All of these frameworks are open
source. Google developed TensorFlow (www.tensorflow.org), and it is known for
having an architecture that allows computation on any CPU or GPU, either on a
desktop, or server, or even on a mobile device. This framework is available in the
Python programming language, and it has C++ API.
Depending on increasing amount of evaluation parameters and the respecting data
size, deep neural network (DNN) node population may reach to millions of levels.
Computational complexity especially raises the training time. There are also
methods for reducing the time and space complexity of the DNNs. Weight pruning
is one of the trends for this purpose. It has been shown that proper pruning of
computational connections (weights) in DNNs may reduce the computational complexity more than 90%, while the accuracy loss remains negligible (Zhang et al.
2018).
7.7 Conclusion
Food security and agricultural resource sustainability problems raise the importance
of optimal management decisions at all levels of agricultural production. Increasing
the crop monitoring accuracy requires more detailed observation data restricted by
its feasibility. Data fusion is an increasing trend for information harvesting, as data
availability and sensor population exponentially increases in accordance with the
computational power. Every crop variety has different growth model parameters that
require various soil and agrometeorological data for their computation. When the
7 Data Fusion in Agricultural Information Systems
135
of parameters is reduced, and this reduces computational cost. It also controls
overfitting.
Deep learning is not only used for the classification of one data type as cropland
cover (Kussul et al. 2017); it also enables fusion of the different types of data layers
either by applying a different kind of decomposed datasets into common fully
connected network as similar to Fig. 7.21 or they can also be merged before the
convolutional decomposition layer.
Residual neural network (ResNets) (Kaiming et al. 2015) is a deep learning
structure that improves the classification performance by utilizing skip connections
or shortcuts to jump over some layers. DenseNets were proposed in 2016, and they
use several parallel skips as an improvement of ResNets. Deep learning provides an
efficient solution in the registration of relevant data input set features that includes
the projection and rotational differences. For example scale-invariant feature transformation (SIFT) algorithm was used to solve especially this feature matching
problem in the past years. Object features are efficiently matched in deep learning
structures without the necessity of additional preprocessing methods (Sachdeva et al.
2017). This property makes deep learning as an efficient candidate, especially for
spatial and spatiotemporal data fusion processes. A widespread implementation
method of the deep learning services is using dedicated frameworks. Tensorflow,
Keras, Pytorch, Caffee, Theano, Apache MXNET, and Microsoft CNTK are widespread deep learning frameworks by the year 2018. All of these frameworks are open
source. Google developed TensorFlow (www.tensorflow.org), and it is known for
having an architecture that allows computation on any CPU or GPU, either on a
desktop, or server, or even on a mobile device. This framework is available in the
Python programming language, and it has C++ API.
Depending on increasing amount of evaluation parameters and the respecting data
size, deep neural network (DNN) node population may reach to millions of levels.
Computational complexity especially raises the training time. There are also
methods for reducing the time and space complexity of the DNNs. Weight pruning
is one of the trends for this purpose. It has been shown that proper pruning of
computational connections (weights) in DNNs may reduce the computational complexity more than 90%, while the accuracy loss remains negligible (Zhang et al.
2018).
7.7 Conclusion
Food security and agricultural resource sustainability problems raise the importance
of optimal management decisions at all levels of agricultural production. Increasing
the crop monitoring accuracy requires more detailed observation data restricted by
its feasibility. Data fusion is an increasing trend for information harvesting, as data
availability and sensor population exponentially increases in accordance with the
computational power. Every crop variety has different growth model parameters that
require various soil and agrometeorological data for their computation. When the
7 Data Fusion in Agricultural Information Systems
135