Spatial Recognition of Citrus Trees, Seedlings, and Gaps: A Deep Learning and Geometric Approach on Satellite-Based RGB Imagery

Global food deprivation is one of the key concerns in recent years leading to the development of several techniques to increase food production. A decisive factor in increasing the overall food production is the optimum use of agricultural land. When planting orchards, proper planning and monitoring is required. Failure to do so can lead to tree gaps or blank spots in the plantation either due to the planting irregularity, or natural reasons like weather, pests, and diseases. Research work has been done on the monitoring of trees and crops using remotely sensed data, but we find limited work on finding gaps in agricultural plantations such as fruit orchards. Contributing to this, we have proposed a deep-learning-based approach that uses satellite data to detect individual tree gaps in citrus orchards in Pakistan. Very High-resolution (VHR) satellite images were collected using the Google Earth application. A custom-trained convolutional neural network is applied to the images. A specialized algorithm based on geometric information computes the location and count of individual trees, seedlings, and tree gaps in the given satellite image. Our proposed technique can detect trees with an F-score of 0.997, seedlings with an F-score of 0.964, and tree gaps with an F-score of 0.951.