Accuracy Assessment Through Contrasting Space-Time Resolutions, a Case Study in Maize Crop Phenology

Physiological crop growth and development assessment is critical for agricultural management and for maintaining long-term sustainability. Understanding the changes in crop developmental phases (herein termed as crop phenology) and their short-term variations (temporal-scale) is essential for production-beneficial management actions such as appropriate irrigation, fertilization, and crop protection. Weekly field survey data gathered from a broad network of regional extension agricultural agents based on their observations in the field is used to calculate current estimations in crop phenology progress. While this is an important source of information, it is a time-consuming, subjected to bias, and labor-intensive process that may not fully represent the fidelity experience of a county or district. Remote sensing could prove useful in this effort by improving overall crop phenology advancement and forecast while also resolving concerns related to bias and missing data. Sensors from current and planned satellite missions are creating a vast universe of information with unprecedented temporal, spectral, and spatial resolutions. High observational frequency and spatial resolution are the most important factors in monitoring crop phenology progress and tracking vegetation dynamics, which is best achieved through data fusion of multi-sensor data streams. To handle large datasets with acceptable trade-off and stability while being computationally efficient, robust classifiers are necessary. Random Forest (RF) has been used for solving tractable remote sensing issues for decades. Its simplicity and high performance, results in superior classification metrics when compared to numerous classifiers, particularly when dealing with large and unbalanced datasets. We compared the output of a RF classifier model utilizing Surface Reflectance data from Planets data fusion product (PF-SR) and ESA's Sentinel 2 sensors, to identify maize phenology in two distinct locations of Kansas (US), during the 2017 growing season. We found that the use of daily data from PF-SR (integrating multi-constellation from PlanetScope, Sentinel-2, and Landsat 8) positively impacted RF classification metrics improving actionable insights and more closely impacting the farming decision support process.