**5. Discussion**

#### *5.1. Effect of Sky Conditions on Albedo Estimation*

We conducted the spectrometer experiment under both clear sky and overcast sky conditions for the two drone field sites. The sky conditions had some effect on the regression statistics. The *R*<sup>2</sup> values for Brooksvale Park were higher if the ground measurements were made under overcast sky conditions (Figure 4), the same conditions as when the UAV mission took place, than if the measurements were made under clear sky conditions. In contrast, *R*<sup>2</sup> values for the Yale Playground were higher under clear sky conditions than under overcast sky conditions (Figure 5), keeping in mind that the UAV mission was conducted there under mostly clear sky conditions. These results indicate that the spectrometer calibration experiment should be conducted under the sky conditions that match those of the UAV experiment.

The Yale Playground flight was conducted at 14:30 on 30 September 2015 when the solar elevation angle was rather low. The long shadows due to the low elevation angle were a source of uncertainty for the albedo estimation. Choosing an appropriate flight time so that the shadow effect is minimized is very important, especially in an urban environment where tall structures are prominent features of the landscape.

The difference between the Landsat- and drone-derived visibleband albedo of the Brooksvale Park were larger than those of the Yale Playground. The main factor contributing to this discrepancy is also the sky condition. The UAV measurement was conducted under overcast conditions at Brooksvale Park, whereas the Landsat measurement was used for clear sky conditions. Generally, surface albedo is higher under cloudy skies than under clear skies [38]. At the Yale Playground, the sky conditions of the UAV experiment matched those of the Landsat observation, resulting in a much better agreemen<sup>t</sup> between the two albedo estimates.

#### *5.2. Uncertainty in Landscape Shortwave Albedo*

Low-cost consumer-grade cameras usually do not contain near-infrared spectral information, and therefore this may cause problems for estimating shortwave band albedo. In this study, shortwave band albedo was determined in a relatively arbitrary way. Brest and Goward [39] assigned weight factors to Landsat visible, near-infrared, and mid-infrared band reflectance, and linearly combined them to estimate the shortwave band albedo for vegetation. Similar to their method, we used the average ratio between the shortwave and visible band albedos of non-vegetation and vegetation ground targets to retrieve the landscape shortwave band albedo. The ratio for clear skies were lower than that for overcast skies for vegetation targets, suggesting that plants preferably absorb more visible radiation under clear skies than under overcast skies. For non-vegetation targets, the ratio did not differ by much between the two different sky conditions. Vegetation ground targets had higher ratios than the non-vegetation targets, due to the high near-infrared spectral reflectance of plant cell structures. The result from Brest's [40] study may provide a useful reference here: They reported that urban downtown and high-density residential neighborhoods in Hartford, Connecticut, USA, had ratios of near-infrared to visible band reflectance between 1.29 and 1.38. Their ratio for evergreen and deciduous forests was between 2.53 and 3.63. From the shortwave and visible band albedos given in their study, we infer that their non-vegetation ratio of shortwave to visible band reflectance is 1.15 to 1.19, which is in good agreemen<sup>t</sup> with our non-vegetation ratios (1.18 to 1.24), and their vegetation (evergreen and deciduous trees) ratio is 1.77 to 2.32, which is much lower than our ratios (3.91 to 6.67). The main reason for the difference is that grass leaves (in this study the vegetation ground targets we chose was grass) have a higher spectral reflectance in the infrared waveband than tree leaves, which can increase shortwave albedo and therefore the ratio between shortwave and visible band reflectance [40].

Three modifications to our method can potentially improve the shortwave albedo accuracy. First, the addition of a light-weight NIRcamera can give a direct reflectance measurement to every ground pixel and therefore avoid the ratio method.

Second, in the current experiments, we were only able to perform a spectrometer measurement of the short grass targets. However, a large portion of the vegetation pixels are trees (Figure 2). Applying the NIR-to-visible band ratio obtained from the grass targets to these tree pixels will cause large errors. This is especially problematic for the Brooksvale Park where some trees started to become senescent (Figure 2a). Division of the vegetation pixels into three separate categories (grass, senescent tree leaf, and green tree leaf) and establishing the NIR to visible band for each category should improve the shortwave albedo accuracy.

Third, we deliberately selected the ground targets to cover a wide range of reflectivity to establish robust calibration curves, but we did not consider variations in BRDF signatures between grasses and trees, and between sands and tarmac. Currently only two training sets (vegetation versus non-vegetation) were used for conversions to shortwave albedo. Additional training sets accounting for BRDF variations may improve our results.
