Hybridizing deep learning algorithms and geostatistical approaches for improved crop yield disaggregation.

Area of Science:

• Agricultural Science
• Geospatial Analysis
• Data Science

Background:

• Accurate crop yield data at fine resolutions is crucial for precision agriculture and food security.
• Existing yield statistics are often aggregated at administrative levels, limiting field-scale applications.
• Disaggregating village-level data to pixel-level resolution presents significant challenges.

Purpose of the Study:

• To develop and validate a hybridized framework for disaggregating village-level crop yield statistics to pixel-level resolution.
• To identify optimal data combinations (soil, weather, satellite imagery) for yield disaggregation.
• To improve the spatial realism and accuracy of crop yield estimation.

Main Methods:

• Integration of deep learning (DL) models with geostatistical residual kriging.
• Evaluation of various data combinations, including Sentinel-1, Sentinel-2, soil, and weather data.
• Application of residual kriging to DL model outputs to correct spatial biases.

Main Results:

• Datasets combining spectral and weather information yielded the best results for disaggregation.
• DL models achieved high numerical accuracy and spatial realism but had structured residuals.
• The hybridized framework reduced RMSE by 35-45%, producing smoother, more realistic pixel-level yield maps.

Conclusions:

• The proposed hybridized framework effectively balances statistical accuracy with spatially realistic yield disaggregation.
• Residual kriging is essential for correcting spatial biases in DL model outputs.
• This study demonstrates a novel approach for village-to-pixel yield disaggregation using integrated data sources.