Efficient analysis and compression of urban green areas in RGB drone imagery using the OSAVI index

Abstract

Green urban area detection is essential for environmental planning; traditional field surveys are laborintensive and time-consuming, making remote sensing (drone and satellite imagery) a powerful alternative. Three main problems are detected by working with these technologies: i) While enabling detailed analysis of lawns and individual trees due to their high spatial resolution, this results in data storage demands. ii) Moreover, the Normalized Difference Vegetation Index (NDVI), the most widely used for analyzing general vegetation, is highly sensitive to soil brightness, making it less suitable for examining urban greenery where bare soil, artificial surfaces, and mixed land covers are common. iii) Additionally, existing tree inventory algorithms in urban or heterogeneous environments remain labor-intensive, as they require annotated training samples to effectively distinguish trees from surrounding features. This study presents high-resolution multispectral and RGB imagery captured by an Unmanned Aerial Vehicle (UAV), the DJI MAVIC 3M, used to measure general vegetation. A masking process based on morphological operations was applied to segment green urban areas in the RGB image, to optimize both image storage size with lossless compression (Deflate & LZW) and traditional tree inventory based on crown detection (DeepForest). The segmentation based on Optimized Soil-Adjusted Index (OSAVI) mask applied in urban areas presents multiple advantages in terms of reduction of storage size due to the increase in homogeneous regions with pixel values sharing identical color characteristics. By using the OSAVI vegetation index as the masking criterion, the dense vegetation (trees) is not affected during the process, preserving its location and color (pixel values) of the original image, excelling current tree inventory algorithms (DeepForest) based on orthoimages without the need to prepare additional training data. Using the OSAVI instead of NDVI outperformed traditional green urban area segmentation, demonstrating 25% more robustness in avoiding saturation caused by Near-Infrared (NIR) reflecting areas. The segmentation performance achieved by OSAVI and morphological operations resulted in: IoU = 0.85 | Dice = 0.91 | Precision = 0.89 | Recall = 0.94 | Accuracy = 0.96. The final storage sizes of the masked RGB images were equal to the percentage of vegetation multiplied by the storage size of the non-masked (original) compressed images, with a Pearson Correlation of 0.98, being the Deflate method superior (bpp = 9) to the LZW method (bpp = 11.40) in terms of storage efficiency. The comparison between the tree inventory on the original RGB scenarios presented a Mean Absolute Error (MAE) = 193.25, and the RGB images masked by OSAVI index MAE = 98.25, 49% better and closer to the real tree inventory, with a Friedman test p-value = 0.046 rejecting the null hypothesis that all methods (Baseline: Precision = 0.33 | Recall = 0.45 | F1 = 0.38 and Proposed: Precision = 0.54 | Recall = 0.53 | F1 = 0.53 ) perform equally.