*4.3. The Sensitivity of Red-Edge Wavelengths to Di*ff*erent Crop Types*

Red-edge refers to the region of the sharp rise in vegetation reflectance between the red and NIR part of the electromagnetic spectrum. It is an important wavelength range that is sensitive to vegetation conditions. This research demonstrated the advantage of adding red-edge spectral bands and indices in the crop type discrimination, which increased the overall accuracy by 3.06% compared with the results only using conventional optical features. In Figure 14, among the selected red-edge features, Band 6 (red-edge 2), Band 5 (red-edge 1), and the NDVIre2n (red-edge index based on Band 6 and Band 8A) exhibited greater importance. As for red-edge indices, 10 out of 11 indices were selected for crop type discrimination (NDVIre3n was discarded by RFI method), with a number of indices showing a good capability to differentiate corn from other crops (Figure 21). This explains why the inclusion of red-edge features significantly improved the mapping accuracy of corn (Figure 13). Apart from corn, some re-edge indices also revealed good separability of tomato (Figure 21c–f,h) and pear (Figure 22). Several red-edge features show good capability to identify chili pepper (Figure 23). These findings reinforce the benefit of using red-edge wavelengths in crop type mapping.

**Figure 21.** Boxplots of selected red-edge indices showing a good capability to distinguish corn from other crops. Each feature is named by its subcategory and acquisition time in the format of "yyyymmdd". (**a**) Boxplots of the NDVIre2n indices derived from the 1 September 2018 acquisition of Sentinel-2; (**b**) boxplots of the NDVIre2n indices derived from the 17 August 2018 acquisition of Sentinel-2; (**c**) boxplots of the CIre indices derived from the 17 August 2018 acquisition of Sentinel-2; (**d**) boxplots of the NDVIre1n indices derived from the 17 August 2018 acquisition of Sentinel-2; (**e**) boxplots of the MSRre indices derived from the 17 August 2018 acquisition of Sentinel-2; (**f**) boxplots of the NDre2 indices derived from the 17 August 2018 acquisition of Sentinel-2; (**g**) boxplots of the NDVIre2 indices derived from the 17 August 2018 acquisition of Sentinel-2; (**h**) boxplots of the MSRren indices derived from the 17 August 2018 acquisition of Sentinel-2; (**i**) boxplots of the NDVIre2 indices derived from the 01 September 2018 acquisition of Sentinel-2.

**Figure 22.** Boxplots of selected red-edge indices revealing a clear distinction between the pear and other crops. Each feature is named by its subcategory and acquisition time in the format of "yyyymmdd". (**a**) Boxplots of the NDVIre1 indices derived from the 9 May 2018 acquisition of Sentinel-2; (**b**) boxplots of the NDVIre1n indices derived from the 6 October 2018 acquisition of Sentinel-2; (**c**) boxplots of the NDVIre2 indices derived from the 9 May 2018 acquisition of Sentinel-2; (**d**) boxplots of the CIre indices derived from the 06 October 2018 acquisition of Sentinel-2.

**Figure 23.** Boxplots of selected red-edge features showing good separability of chili pepper. Each feature is named by its subcategory and acquisition time in the format of "yyyymmdd". (**a**) Boxplots of the band 6 values from the 17 August 2018 acquisition of Sentinel-2; (**b**) boxplots of the band 6 values from the 1 September 2018 acquisition of Sentinel-2; (**c**) boxplots of the band 7 values from the 17 August 2018 acquisition of Sentinel-2.

From the top six scoring features for crop type mapping (Figure 8) and the accumulated importance score of each subcategory (Figure 14b), it can be inferred that the most important red-edge band is Band 6, and the most useful red-edge indices is the NDVIre2n (calculated using Band 6 and Band 8A (NIR narrow)). The top six features emphasized the significance of three red-edge features, i.e., Band 6 (red-edge 2), Band 7 (red-edge 3), and NDVIre2n (red-edge index based on Band 6). The accumulated importance ranking (Figure 14b) indicated the contribution of Band 6, Band 5, and NDVIre2n. It has been reported that the red-edge close to red wavelengths (Band 5) is mainly related to the difference in chlorophyll content, while the red-edge close to NIR (Band 7) is usually correlated to variations in the leaf structure [6]. The above results suggest that the separability of red-edge features lies in both the leaf structure and chlorophyll content of different crop species.

In a nutshell, SAR features distinguish crops based on the structural and geometrical arrangements of plants or changes of the crop field surface, while optical features rely on the multi-spectral information to differentiate crops. In experiments, the red-edge wavelengths exhibited good separability between different crop types, due to their sensitivity to the variations in chlorophyll content and leaf structure. Therefore, the combination of SAR and optical integrated the physical and spectral characteristics of crops, improving the performance of crop classification.

With the development of deep learning technology, powerful convolutional networks, such as U-net [39], can be explored for crop type mapping in future work. U-net has the potential to support multi-dimensional input variables. By implementing U-net with multi-temporal, multi-sensor features, it is expected that a higher level of accuracy can be achieved in crop type identification.
