time. Each TARBIL monitoring station has two or three cameras for crop development and phenological observations. Their images are used for ground reference
data together with official records from farm registry system for crop pattern
classification.
Crop yield maps are extracted by using these two maps as yield efficiency
(Fig. 7.9) and crop area intensity (Fig. 7.10). This process is applied separately for
irrigated and dry farming conditions. Overall crop yield of the province is then
computed by using plantation intensity and yield efficiencies both for irrigated and
dry farming conditions. Harvesting data records from 275 harvesting machines
having GPS-coordinated monitoring device at the main cereal basins of Turkey are
used for the additional calibration of the model in 2015. Additional statistical
calibration improves estimation (nowcasting) accuracy because of several
nonnatural factors, including regional change in harvesting methods and seed variety
preferences.
Overfitting is one of the critical issues in machine learning or statistical regression
model-based estimations. It refers to when a model is so tuned to the training
examples that it is not able to generalize well for the validation or test sets. A
symptom of overfitting is having a model that gets half of the percentage of the test
data that perfectly fits to the training data.
The ratio between the number of parameters and the acquired dataset size is
important in avoiding the overfitting depending on the nonlinearity of the system.
For this reason, regression models that use less amount of parameters should be
preferred for equal or close correlation rates.
7.4 Neural Networks for Data Fusion
The main advantage of the neural networks is that they can learn from past experiences, which allows them to learn and adapt to changes. Neural network
(NN) models represent a wide class of flexible nonlinear models which are used
for such purposes as classification, pattern recognition, clustering, anomaly detection
and forecasting. It has been shown that properly defined NN models achieve more
reliable predictions than the conventional regression methods (Sarmadian and
Mehrjardi 2008). In addition, the usage of NN provides better solutions when they
are applied to the complex systems that may be poorly understood by the traditional
analytical methods (Tokar and Markus 2000). NNs may be described as a network of
interconnected neurons (nodes). Each neuron consists of several input nodes and an
output node. The problem specifies the requried number of neurons at the input
layer, hidden layer, and the output layer of the neural network. Commonly used
simple neuron models compute the output of the neurons based on the weighted sum
of all its inputs according to an activation function. Arranging the data has a
significant impact on the obtained results of the trained network. There is not a
unique neural network structure yet that can fit to every kind of pattern recognition problem. For this reason, evaluation of the priory information about time
122
B. Üstündağ
data together with official records from farm registry system for crop pattern
classification.
Crop yield maps are extracted by using these two maps as yield efficiency
(Fig. 7.9) and crop area intensity (Fig. 7.10). This process is applied separately for
irrigated and dry farming conditions. Overall crop yield of the province is then
computed by using plantation intensity and yield efficiencies both for irrigated and
dry farming conditions. Harvesting data records from 275 harvesting machines
having GPS-coordinated monitoring device at the main cereal basins of Turkey are
used for the additional calibration of the model in 2015. Additional statistical
calibration improves estimation (nowcasting) accuracy because of several
nonnatural factors, including regional change in harvesting methods and seed variety
preferences.
Overfitting is one of the critical issues in machine learning or statistical regression
model-based estimations. It refers to when a model is so tuned to the training
examples that it is not able to generalize well for the validation or test sets. A
symptom of overfitting is having a model that gets half of the percentage of the test
data that perfectly fits to the training data.
The ratio between the number of parameters and the acquired dataset size is
important in avoiding the overfitting depending on the nonlinearity of the system.
For this reason, regression models that use less amount of parameters should be
preferred for equal or close correlation rates.
7.4 Neural Networks for Data Fusion
The main advantage of the neural networks is that they can learn from past experiences, which allows them to learn and adapt to changes. Neural network
(NN) models represent a wide class of flexible nonlinear models which are used
for such purposes as classification, pattern recognition, clustering, anomaly detection
and forecasting. It has been shown that properly defined NN models achieve more
reliable predictions than the conventional regression methods (Sarmadian and
Mehrjardi 2008). In addition, the usage of NN provides better solutions when they
are applied to the complex systems that may be poorly understood by the traditional
analytical methods (Tokar and Markus 2000). NNs may be described as a network of
interconnected neurons (nodes). Each neuron consists of several input nodes and an
output node. The problem specifies the requried number of neurons at the input
layer, hidden layer, and the output layer of the neural network. Commonly used
simple neuron models compute the output of the neurons based on the weighted sum
of all its inputs according to an activation function. Arranging the data has a
significant impact on the obtained results of the trained network. There is not a
unique neural network structure yet that can fit to every kind of pattern recognition problem. For this reason, evaluation of the priory information about time
122
B. Üstündağ