Fusing Cubesat and Landsat 8 data for near-daily mapping of leaf area index at 3 m resolution

Constellations of small cubesats are emerging as a relatively inexpensive observational resource with the potential to overcome spatio-temporal constraints of traditional single-sensor satellite missions. With more than 130 compact 3U (i.e., 10 x 10 x 30 cm) cubesats currently in orbit, the company "Planet" has realized near-daily image capture in RGB and the near-infrared (NIR) at 3 m resolution for every location on the earth. However cross-sensor inconsistencies can be a limiting factor, which result from relatively low signal-to-noise ratios, varying overpass times, and sensor-specific spectral response functions. In addition, the sensor radiometric information content is more limited compared to conventional satellite systems such as Landsat. In this study, a synergistic machine-learning framework utilizing Planet, Landsat 8, and MODIS data is developed to produce Landsat 8 consistent LAI with a factor of 10 increase in spatial resolution and a daily observing potential, globally. The Cubist machine-learning technique is used to establish scene-specific links between scale-consistent cubesat RGB+NIR imagery and Landsat 8 LAI. The scheme implements a novel LAI target sampling technique for model training purposes, which accounts for changes in cover conditions over the cubesat and Landsat acquisition timespans. Results over an agricultural region in Saudi Arabia highlight the utility of the approach for detecting high frequency (i.e., near-daily) and fine-scale (i.e., 3 m) intra-field dynamics in LAI with demonstrated potential for timely identification of developing crop risks. The framework maximizes the utility of ultra-high resolution cubesat data for agricultural management and resource efficiency optimization at the precision scale.