I am working on a project to develop an AI-powered forestry management system for monitoring and analyzing tree populations, focusing on counting trees and identifying endangered or indigenous species. The project aims to use satellite imagery from sources like Google Earth and drone footage.

I have implemented a basic U-Net model for image segmentation to detect and count trees in satellite images. However, the current model does not perform accurately, missing many trees and incorrectly counting others. I would like to use a pretrained model to improve the accuracy and efficiency of tree detection and counting.

My Current Workflow:
Data Acquisition: Collecting satellite images from Google Earth.
Preprocessing: Normalizing and enhancing images.
Image Segmentation: Using U-Net for initial tree crown segmentation.
Tree Counting: Applying contour detection to count trees.
Species Identification: Aiming to classify tree species based on segmented regions (not yet implemented).
Problems I’m Facing:
Accuracy: The model misses many trees and gives incorrect counts.
Pretrained Models: Difficulty in integrating a high-quality pretrained model to improve segmentation results.
Data Sources: Uncertainty about where to find or how to create a high-quality training dataset for better results.

Questions:
Pretrained Models: What are the best pretrained models available for tree detection in satellite imagery, and how can I effectively integrate them into my project?
Dataset Recommendations: Where can I find high-quality datasets for tree detection and species identification in satellite or aerial imagery?
Improving Accuracy: What techniques or additional preprocessing steps can I take to improve the accuracy of tree detection and counting?
Additional Information:
I am using Python with libraries such as TensorFlow, Keras, and OpenCV.
The project is crucial for my graduation, and any detailed guidance or references to relevant tutorials or resources would be highly appreciated.

I attempted to use a pretrained U-Net model from the segmentation_models library, but faced errors related to layer shapes and input dimensions.
I’ve explored various datasets but am unsure which ones would be most effective for training a robust model.