A deep learning-based framework for object recognition in ecological environments with dense focal loss and occlusion
Muhammad Munir Afsar,Asim Dilawar Bakhshi,Ejaz Hussain,Javed Iqbal
DOI: https://doi.org/10.1007/s00521-024-09582-5
2024-03-07
Neural Computing and Applications
Abstract:In precision agricultural analysis, the remote sensing of geospatial data holds substantial potential for multi-purpose crop surveys, targeting automatic crop area delineation, health monitoring, and yield estimation. Advanced remote sensing methods, when integrated with machine learning techniques, have significantly advanced agricultural analyses. This study introduces a three-tiered framework. Firstly, orchard areas are delineated using the ESA Sentinel-2 multispectral instrument (MSI) at 10 m resolution, employing the normalized differential vegetation index (NDVI). In the second stage, mango tree canopies are detected from hand-annotated true color composite imagery using two variants of a convolutional neural network. The first variant, CanopyNet-1, is built directly over RetinaNet foundational layers, achieving a mean average precision (mAP) of 0.79, a precision of 0.80, and a recall of 0.76. The second variant, CanopyNet-2, builds upon DeepForest, a generalized tree canopy trained model, also using RetinaNet at its base. CanopyNet-2 demonstrates superior performance, achieving a mAP of 0.83, a precision of 0.98, and a recall of 0.96, notably surpassing conventional models such as YOLOv5 and Faster R-CNN. Lastly, the health of the orchard is characterized using 3-m-resolution multispectral imagery. Cumulatively, our framework, with its tiered approach, exhibits high accuracy in both tree canopy delineation and health characterization, suggesting it as a comprehensive solution for large-scale orchard monitoring and yield optimization.
computer science, artificial intelligence