classifications of crop and land cover if the application is not across sensors and in
similar seasons (Vicenteserrano et al. 2008; Hansen and Loveland 2012). Topographic normalization is needed if classification is applied across sensors and
seasons (Lu et al. 2008).
Clouds and hazes block the view of the ground in optical remote sensing
(Whitcraft et al. 2015). They need to mask out and mark in quality layer for
further processing. There are many detection algorithms available to detect and
mask the cloud pixels in optical remotely sensed data (Lyapustin et al. 2008;
Hulley and Hook 2008; Zhu and Woodcock 2012).
3. Feature extraction and feature selection: Feature extraction is the process to
derive values and features from remotely sensed data. The examples of derived
features are textural features, statistical features, Discrete Wavelet Transform
(DWT)-based feature, and Discrete Cosine Transform (DCT)-based feature
(Badhwar et al. 1982; Lei et al. 2008; Anami et al. 2011; Ul Qayyum et al.
2013). Feature selection is the process of selecting the proper features for crop
classification. The use of proper features for crop classifications depends on crop
types, classification algorithms, sensors, and seasons. Too many features may
lead to degraded crop classification (Lu and Weng 2007; Löw et al. 2013). There
are many feature selection methods, e.g., exhaustive search by recursion (ESR),
isolated independent search (ISS), and sequence-dependent search (SDS) (Peddle
and Ferguson 2002). The proper feature extraction and feature selection can
improve the accuracy of crop classification (Löw et al. 2013).
4. Classification: Many classification algorithms have been applied in cropland
classification from remotely sensed data. Both unsupervised and supervised
classification algorithms have been applied. Unsupervised classification algorithms leave the label assignment in post-processing, while supervised classification completes the label assignment during the classification. The hybrid
approach of using both unsupervised and supervised classifications is possible,
while unsupervised classifier is applied first to generate clusters and supervised
classifier is applied on clusters. The unsupervised stage can be seen as a variation
of the feature extraction. Table 10.2 shows some of the commonly used classification algorithms. Ensemble classification (e.g., random forest) and deep learning (e.g., convolutional neural network (CNN)) have been gaining popularity
lately due to its improved accuracy.
5. Post-processing: Filters are normally good at removing the “salt-and-pepper”
effect where misclassified, isolated pixels are eliminated with the assumption of
single crop in a field or a continuous segment of field (Lu and Weng 2007).
Mosaics of cropland from the classification results from time series of remotely
sensed data with ancillary data would improve the classification accuracy. For
example, rule-based reasoning may be applied to mask out built-up areas and
forest areas which are relatively unchanged over time. The planting difference
180
E. G. Yu and Z. Yang
similar seasons (Vicenteserrano et al. 2008; Hansen and Loveland 2012). Topographic normalization is needed if classification is applied across sensors and
seasons (Lu et al. 2008).
Clouds and hazes block the view of the ground in optical remote sensing
(Whitcraft et al. 2015). They need to mask out and mark in quality layer for
further processing. There are many detection algorithms available to detect and
mask the cloud pixels in optical remotely sensed data (Lyapustin et al. 2008;
Hulley and Hook 2008; Zhu and Woodcock 2012).
3. Feature extraction and feature selection: Feature extraction is the process to
derive values and features from remotely sensed data. The examples of derived
features are textural features, statistical features, Discrete Wavelet Transform
(DWT)-based feature, and Discrete Cosine Transform (DCT)-based feature
(Badhwar et al. 1982; Lei et al. 2008; Anami et al. 2011; Ul Qayyum et al.
2013). Feature selection is the process of selecting the proper features for crop
classification. The use of proper features for crop classifications depends on crop
types, classification algorithms, sensors, and seasons. Too many features may
lead to degraded crop classification (Lu and Weng 2007; Löw et al. 2013). There
are many feature selection methods, e.g., exhaustive search by recursion (ESR),
isolated independent search (ISS), and sequence-dependent search (SDS) (Peddle
and Ferguson 2002). The proper feature extraction and feature selection can
improve the accuracy of crop classification (Löw et al. 2013).
4. Classification: Many classification algorithms have been applied in cropland
classification from remotely sensed data. Both unsupervised and supervised
classification algorithms have been applied. Unsupervised classification algorithms leave the label assignment in post-processing, while supervised classification completes the label assignment during the classification. The hybrid
approach of using both unsupervised and supervised classifications is possible,
while unsupervised classifier is applied first to generate clusters and supervised
classifier is applied on clusters. The unsupervised stage can be seen as a variation
of the feature extraction. Table 10.2 shows some of the commonly used classification algorithms. Ensemble classification (e.g., random forest) and deep learning (e.g., convolutional neural network (CNN)) have been gaining popularity
lately due to its improved accuracy.
5. Post-processing: Filters are normally good at removing the “salt-and-pepper”
effect where misclassified, isolated pixels are eliminated with the assumption of
single crop in a field or a continuous segment of field (Lu and Weng 2007).
Mosaics of cropland from the classification results from time series of remotely
sensed data with ancillary data would improve the classification accuracy. For
example, rule-based reasoning may be applied to mask out built-up areas and
forest areas which are relatively unchanged over time. The planting difference
180
E. G. Yu and Z. Yang