Monitoring agricultural behavior under climate change with cloud computing and satellite imagery
- Author(s): Zhang, Minghui;
- Advisor(s): Thompson, Sally;
- Chow, Fotini K
- et al.
Understanding agricultural productivity under climate change is critical to helping global food systems tackle impending challenges in supply and demand. Crucially, this requires us to understand planting dates: by controlling the yield and cropping intensity of rainfed agriculture, planting dates could serve as a major adaptation strategy under climate pressure. Currently, the lack of spatiotemporally resolved crop timing information makes it difficult to produce insights into planting behavior, which is the result of complex human decisions made under varying socio-economic and climatic contexts. This data gap hinders our ability to understand how farmers will adapt to climate change by shifting planting dates, and may negatively impact the accuracy of yield predictions. This dissertation addresses this data gap by introducing a scalable method to estimate planting dates. I apply this method to generate insights into historical and future planting dates of soy (Glycine max) in the heavily agricultural state of Mato Grosso, Brazil (MT). My work begins with a remote sensing-based method to estimate field-scale (500 m) planting and harvest dates over large (100,000 km2), and therefore computationally challenging areas, with sparse ground truth information. The method pairs (1) a timeseries analysis algorithm for MODIS imagery, implementable on the cloud computing platform Google Earth Engine (GEE), to extract 500 m phenological milestones and (2) proxy ground truth data based on Planet Labs imagery to relate phenological milestones to observed planting and harvest dates. Next, I build a statistical model of satellite-estimated planting dates as a function of the wet season onset. This model reveals several novel insights about agricultural behavior in Mato Grosso. First, traditional climatological definitions of wet season onset are less correlated to observed planting dates than alternative, easily observable definitions based on rainfall frequency, highlighting the need to explore farmer-relevant definitions of climate. Second, planting dates' sensitivity to wet season onset varies dramatically among fields and with cropping intensity; this heterogeneous response produces a wide range of planting dates that are rarely included in crop models. Finally, a trend toward earlier planting dates, independent of wet season onset, reveals nonstationary behavior that cannot be captured with the historical survey data used by some yield prediction efforts. My findings suggest that under RCP 8.5 climate conditions, climatic windows will constrain planting dates relative to agronomically preferred times, and the feasibility of double cropping will be endangered for vulnerable portions of Mato Grosso. Both delayed planting dates and loss of double cropping suitability are problematic for an economy that largely depends on agribusiness and that is central to international soy supply. While Mato Grosso is the focus of this dissertation, the methods developed here lay the groundwork for similar studies globally. By introducing a scalable method to close the information gap on planting dates and generating new insights into the planting dates of tropical rainfed crops, this work provides a foundation for investigating planting date behavior and climate change adaptation in vulnerable, data-scarce agricultural regions worldwide.