Geospatial assessment of agroforestry land use systems using very high-resolution satellite images and artificial intelligence

In a vast agrarian country like India, multifunctional agroforestry land use systems play significant role. Reliable scientific database on agroforestry systems are essential to implement National Agroforestry Policy of Government of India. This study is a flagship initiative to develop a scalable geospatial approach for assessment of agroforestry resources using sub-meter satellite images and Deep Learning (DL) techniques. A reliable methodology was developed to quantify tree components in Indian agroforestry systems, using U-Net based DL architecture pre-trained on ResNet 34 backbone. 12,628 labelled training samples and in-season ground truth information from 1,767 locations, representing diverse agro-ecological regimes from 6 study areas were utilised for DL model development. Model accuracy was estimated as 93.72 percent, underscoring its robustness to extract major tree components like individual trees on farmland, linear and block plantations. DL outputs were integrated with harmonized Land Use Land Cover maps at 1:10,000 scale, to arrive at integrated agroforestry land use map with 10-15 classes, with an overall accuracy of 86.5 percent and kappa coefficient of 0.847. This is the first detailed study in India, adopting AI based technique for classifying tree components using sub-meter images within large geographic extent of 25,501 km2. It is a major step towards establishing improved geospatial procedure for scientific assessment of agroforestry systems to meet India's commitments to UNFCCC and Intended Nationally Determined Contributions at COP21 in Paris on climate and environment under International Charters. This robust, DL based approach offers potential applications in ecological research, upliftment of agrarian community, sustainability and environmental policy planning.