Crop loss identification at field parcel scale using satellite remote sensing and machine learning.

Area of Science:

• Agricultural remote sensing
• Machine learning applications in agriculture
• Environmental monitoring

Background:

• Identifying crop loss at the field parcel scale using satellite imagery is difficult due to multiple causal factors, limited reference data, and varied definitions of crop loss.
• Accurate crop loss assessment is crucial for agricultural management, insurance, and government policy.
• Existing methods often lack the scale and precision required for detailed agricultural monitoring.

Purpose of the Study:

• To investigate the feasibility of using satellite imagery to train machine learning models for classifying agricultural field parcels with and without crop loss.
• To evaluate the performance of different machine learning models in a large-scale crop loss identification task.
• To establish a benchmark for machine learning-based crop loss classification at the field parcel scale.

Main Methods:

• Utilized a comprehensive dataset of approximately 1.4 million Finnish field parcels from 2000-2015, including crop loss information from the Finnish Food Authority (FFA).
• Integrated Normalized Difference Vegetation Index (NDVI) data derived from Landsat 7 imagery, addressing significant data gaps (over 80% missing values).
• Trained and compared four machine learning models, focusing on Random Forest with mean imputation and missing value indicators, to classify crop loss.

Main Results:

• The Random Forest model, employing mean imputation and missing value indicators, achieved an average Area Under the ROC Curve (AUC) of 0.688±0.059 across 16 years.
• The model demonstrated consistent performance in identifying new crop-loss fields, with AUC values ranging from 0.602 to 0.795.
• Despite extremely noisy data and missing values, the study successfully demonstrated the potential of machine learning for crop loss classification.

Conclusions:

• Machine learning models, particularly Random Forest, can effectively classify crop loss at the field parcel scale using satellite data, even with substantial data limitations.
• The developed classification framework and models offer practical applications for government agencies and insurance companies in verifying crop loss claims and enhancing agricultural monitoring.
• This study represents a significant contribution to large-scale benchmark research on machine learning for agricultural field parcel analysis and crop loss detection.