Estimation of different data compositions for early-season crop type classification

Area of Science:

• Agricultural Remote Sensing
• Geospatial Analysis
• Vegetation Monitoring

Background:

• Accurate crop type distribution maps are crucial for yield estimation and production forecasting.
• Time series remote sensing data, particularly Normalized Difference Vegetation Index (NDVI), are essential for crop type mapping.
• Image composition strategies are used to enhance time series data quality, but optimal periods remain unclear.

Purpose of the Study:

• To evaluate the impact of different NDVI time series composition periods (daily, 8-day, 16-day, 32-day) on crop classification accuracy.
• To determine the optimal composition period for crop type mapping using fused remote sensing data.
• To assess the early identification of crops like cotton using various composition strategies.

Main Methods:

• Fused daily 30m NDVI time series from multiple sensors (MODIS, Landsat, Gaofen, HJ).
• Composited NDVI time series using four strategies: daily, 8-day, 16-day, and 32-day.
• Employed Random Forest classifier for crop type identification in Xinjiang, China.

Main Results:

• Daily NDVI time series yielded the highest crop classification accuracies, with overall accuracies reaching 86.13% (Bole) and 91.89% (Luntai).
• Cotton was identified significantly earlier (40–60 days prior to harvest) using daily, 8-day, and 16-day compositions with high producer's and user's accuracies (>85%).
• While 8-day, 16-day, and 32-day compositions showed similar saturated accuracies, the 8-day and 16-day achieved them earlier than the 32-day composition.

Conclusions:

• Daily NDVI time series are superior for crop classification accuracy and early crop identification.
• The 16-day NDVI composition is a viable alternative when daily data acquisition is not feasible, offering a balance between accuracy and temporal resolution.
• Optimizing NDVI composition periods is critical for improving the reliability of crop type distribution maps derived from remote sensing data.