Evaluation of CubeSat-retrieved LAI
Independent assessment of the LAI values derived from the CubeSat-based approach was performed against the field-derived LAI measurements. As can be seen in Figure 2, there was a strong linear relationship between the field and CubeSat-based LAI measurements, producing coefficient of determination (R2) values > 0.94. Interestingly, the rRMSE for US-Ne3 (12.74) was significantly lower than those from the two irrigated fields (16.54 and 17.96; Figure 2), suggesting more accurate LAI estimates within the rain-fed field. LAI values ≤ 5 generally followed a linear relationship, whereas the field-based LAI values > 5 were in most cases underestimated by the CubeSat-based approach; a common characteristic identified in other studies using satellite imagery for estimating LAI [e.g. 43,44]. While the laboratory-based measurements of LAI represented the green leaf cover, the CubeSat image data provide a top-down view, which may be affected by the leaf angle distribution, clumping effects, senescent leaves, plant stems, and any exposed soil background. Hence, variations between the two estimates can be expected, especially due to changing soil color and texture or the presence of weeds [45]. Canopy bi-directional reflectance, including that associated with canopy shadowing, may also affect image-derived vegetation structural measurements such as LAI [46]. The random forest regression approach applied to estimate LAI at the CubeSat resolution used vegetation indices known to be sensitive to LAI [11]. Previous studies have found that vegetation indices can saturate at high LAI values, i.e. become insensitive to variations in LAI values above a certain threshold, due to the horizontal leaf stratification [47], which provides a likely explanation of the CubeSat-based LAI underestimation in fields US-Ne1 and 2 from the beginning of July to mid-August depicted in Figure 2. As the rain-fed US-Ne3 field generally had LAI values < 5, saturation effects were therefore not a strong consideration. Sadeh et al. [28] found fused Sentinel-2 and CubeSat data derived LAI values > 3 of wheat fields to saturate, requiring model adjustments. After their model adjustment, R2 ranged from 0.27-0.92, with RMSEs of 0.35-0.63 depending on the vegetation indices used. Kimm et al. [18] produced LAI approximations for corn with an R2 of 0.76 and a RMSE of 1.12 based on CubeSat data. As such, our retrieval accuracies are within or better than those reported by both Kimm et al. [18] and Sadeh et al. [28], and are similar or exceed LAI estimates from Sentinel-2 [e.g. 48,49] and Landsat [e.g. 50,51,52].
Spatiotemporal LAI dynamics drive field insights
The precise application of inputs such as irrigation, fertilizer, herbicides, and pesticides at the right time, rate and place within a field [53] requires spatiotemporally dense data delivered with a low latency. Having a spatial resolution sufficient to identify intra-field variability, while obtaining daily temporal updates on the progression of plant growth, would provide a level of actionable intelligence that farmers could exploit to tailor management decisions for yield optimization [54,55]. High spatiotemporal resolution data also allows the consequences of management decisions and interventions to be assessed. For instance, fields US-Ne1 and Ne2 were planted on April 19 and 23, 2019, respectively. The 4-day time gap between planting of the two fields can be observed in Figure 3, with the progression of LAI values in US-Ne2 consistently lagging those of US-Ne1. The non-irrigated US-Ne3 field that was planted on April 24 appears to be approximately 5 days behind the irrigated US-Ne2 field, even though there was only 1 day difference in planting. The relative temporal lag in LAI development between the three fields remained throughout the vegetative stage (Figure 3).
Intra-field variability in LAI is evident across all three fields, with underperforming darker patches easily recognizable during the greening period displayed in Figure 3 (June 13-July 1), and even at the very beginning of plant growth (June 13). While these are most likely the imprint of underlying soil characteristics, irrigation issues, or inconsistencies during planting [56], the spatially dense CubeSat retrievals provide detail that would likely be unresolved using coarser scale satellite platforms such as Landsat [11] and perhaps even Sentinel-2 [57]. The full resolution (i.e., 3.125 m) of the CubeSat LAI maps permits an improved delineation of field sections showing optimal or suboptimal growth performance. The advantage of the daily image sequence is particularly pronounced given the rapid crop development during the vegetative stage of maize. It is noticeable at the time of peak LAI values (July 14) that the underperforming patches during the vegetative stage tend to dissolve and reach LAI values similar to the remaining parts of the fields, particularly in the US-Ne2 field, which received two lots of 31 mm irrigation between July 1-14.
Identifying different growth stages of maize is a key component for precision insights, as plants have different requirements at different phenological stages [7]. To investigate the relation of CubeSat-derived LAI with key phenological stages of maize, box-and-whisker plots were produced to show the range of LAI pixel values within each field at specific field-identified vegetative and reproductive growth stages (Figure 4). At the early vegetative stage, representing the second leaf collar (V2) at the end of May, CubeSat-observed LAI values can be seen to increase. Subsequently, rapid LAI increases occur in response to plant growth and the production of additional leaf collars (e.g. V6 and V11). The vegetative stages see high nutrient uptakes and rapid plant growth, which relies on moisture, temperature and light interception [25]. Maximum LAI was reached around the time of tasseling (VT, at the beginning of July) and silking (R1, in mid-July), which is the first reproductive stage of maize, coinciding with plant growth that shifts towards pollination and kernel formation. Plant stress (nutrient and moisture deficiencies) prior to tasseling and during pollination can greatly affect pollination and yield, emphasizing the need for careful management inputs in relation to maize phenology [58]. Based on Figure 4, the tasseling stage was clearly identified as the peak, with a subsequent plateau in LAI, demonstrating the capacity to use daily LAI information for fertilizer and irrigation scheduling.
During the first four reproductive stages of kernel development, i.e. silking (R1), blister (R2), milk (R3) and dough (R4), no distinct variation in LAI was observed, despite the continuing increase in reproductive biomass. During kernel development, moisture stress, coupled with high temperatures, nutrient deficiency, disease or insect attack can significantly reduce kernel size and yield [58]. Hence, sudden drops in LAI values during the initial reproductive stages may be an indication of biotic or abiotic stress, requiring urgent management intervention. However, the spectral saturation at high LAI values may have contributed to the similarity in the observed LAI from the tasseling stage and during the first four reproductive stages, which may conceal fluctuations in LAI values above 5, especially in the irrigated fields. The LAI of maize fields depends on the planting density, plant variety, rainfall, temperature, sunlight, availability of nutrients and overall crop growth [59]. In some parts of the world, the LAI of maize is generally below 5 [43,60,61,62], which would eliminate saturation issues. Towards the dent stage (R5) at the end of August and beginning of September, the number of green leaves started decreasing, while the number of dead leaves increased, resulting in a distinct decrease in LAI. According to O’Keeffe [58], the last irrigation should generally occur when the grain is well dented (a couple of weeks before maturity) to reach maximum yield, emphasizing the benefits of daily LAI information, as the R5 stage could clearly be identified in Figure 4 for all three fields. The senescence of green leaves progressed rapidly towards the second half of September, when the plants reached physiological maturity (R6) and maximum biomass accumulation. Approximately four weeks prior to harvest (beginning of November), green LAI values approached zero, as all leaves had dried at that stage. As the yellow maize was used for cattle feed, it is often common practice to postpone harvest until all plants are fully dried for silo storage, hence highlighting another advantage of daily LAI information for harvest scheduling.
Not surprisingly, given the increased potential for plant stress due to water deficit, the rain-fed field (US-Ne3) displayed a larger range of LAI values compared to US-Ne1 and Ne2 in the hotter and drier months of July, August and September (Figure 4). US-Ne3 also experienced a slightly delayed greening stage and started to senesce, reflected by a drop in green LAI at the end of August, and before both US-Ne1 and Ne2. More generally, US-Ne3 displayed lower LAI values than both US-Ne1 and Ne2, indicating the increased production of biomass due to the irrigation of the other two maize fields. In fact, field measurements of biomass for the three fields at the time of plant physiological maturity (R6) showed a difference of approximately 7,000 kg/ha between the irrigated (25,000 kg/ha) and the rain-fed (18,000 kg/ha) fields. Larger production differences between irrigated and rain-fed maize are likely to be more distinct in drier years, where the timing and amounts of irrigation can significantly impact end-of-season yield [63].