Satellite-Based Field Boundaries Detection in India

Accurate field boundary detection (FBD) is crucial for advancing climate-smart agriculture, especially in regions with small, irregular farms like India, where quantification of management impacts is challenging. This study uses the Fractal ResUNet model, an advanced convolutional neural network for FBD, tested on diverse agroclimatic regions using multi-source satellite data. The Fractal ResUNet encoder-decoder architecture excels in learning complex spatial-contextual features from images, enabling precise identification of sparse field contours. Training data ($\mathbf{n}=\mathbf{1 0, 2 8 3}$) were generated using an initial, fine-tuned Segment Anything foundational Model (SAM) combined with high resolution multispectral data (RGB google base maps) implemented in the MXNet framework. Outcomes include extent, boundary, and distance masks over multiple epochs to minimize loss. The key parameters for the model included an initial filter count of 32, a depth of six layers, and a single-class output for field boundary identification. The training was conducted using a Tesla V100 GPU (that contains 5,120 CUDA Cores) GPU with the Adam optimizer and a multi-task loss function to evaluate the model performance. Post-processing involved Higra segmentation and smooth blending to refine boundary delineation. Withheld data were used to validate configurations and performance with F1 and IoU metrics. Results indicate the robust performance of Haryana and Punjab FBD’s using google basemap with F1 scores of 0.86 to 0.84 respectively. In this research application we also evaluated the models using MCC and IoU and discuss deploying Fractal ResUNet in challenging small holder landscapes. The FBD outcomes in this study help quantify management impacts of climate smart practices such as seeding, irrigation, and residue. Further refinement and multi-source fusion are anticipated to further improve the model accuracy and adaptability across different regions.