Farmland quality assessment using deep fully convolutional neural networks.

Area of Science:

• Agricultural Science
• Remote Sensing
• Computer Science

Background:

• Farmland assessment is crucial for food security but traditional methods are slow and costly.
• Deep learning has shown great promise in image recognition and semantic understanding applications.
• Existing farmland evaluation techniques often lack efficiency and scalability.

Purpose of the Study:

• To evaluate farmland grades using deep learning models.
• To compare the performance of fully convolutional networks (FCN) with traditional machine learning methods.
• To explore the use of remote sensing data for efficient farmland assessment.

Main Methods:

• Utilized fully convolutional networks (FCN) as the deep learning model.
• Employed Normalized Difference Vegetation Index (NDVI) derived from Landsat images as input data.
• Trained the model on the China National Cultivated Land Grade Database and applied image segmentation for refinement.

Main Results:

• The FCN achieved an overall F1 score of 0.719 for farmland grade prediction.
• Specific F1 scores for non-farmland, level I, II, III, and IV farmland were 0.909, 0.590, 0.740, 0.642, and 0.023, respectively.
• Combining FCN with image segmentation improved accuracy, reduced noise, and enhanced edge realism compared to conventional machine learning.

Conclusions:

• Deep learning, particularly FCN, provides an acceptable and efficient method for farmland assessment using remote sensing NDVI data.
• This approach serves as a valuable supplement to traditional methods, significantly saving time and cost.
• The study demonstrates the potential of FCN for accurate and scalable agricultural land evaluation without the need for fieldwork.