The growing need for accurate and efficient land cover classification and disturbance mapping has driven the development of advanced methodologies and technologies. One significant challenge faced by researchers and scientists is the collection and use of high-quality training data. This hurdle is especially noticeable for large area classifications, where training data must be defined with minimal error and at higher spatial resolution than the satellite data being classified or validated. To address this need, the use of commercial high-resolution data (less than 10 meters) enables accurate training data generation for applications using 30-meter Landsat and 10 to 20-meter Sentinel-2 data. Implementing deep learning and active learning techniques is essential to leveraging this high-resolution data effectively.
1. Initial Training
Initially, the development of a U-Net model utilizes two primary approaches to obtain training data. First, researchers train U-Net models on coarser resolution burned and unburned validation data. This data originates from a globally distributed set of manually labeled 30-meter Landsat images, providing a diverse and extensive training dataset for the model. Alternatively, high-resolution PlanetScope images of forest environments are manually annotated to indicate tree and no-tree classes. This manual annotation process ensures precise training data, reflecting the intricate variations present in natural landscapes.
By leveraging these two approaches, researchers develop an initial U-Net model capable of classifying land cover types and disturbances with a reasonable level of accuracy. The higher resolution of the PlanetScope images allows for more detailed mapping compared to the coarser resolution Landsat images. The combination of these data sources produces a robust starting point for the model, ready for further refinement and accuracy improvement through the active learning process.
2. Preliminary Classification
Once the initial U-Net model is trained, the next step involves applying it to classify a small number of unlabeled PlanetScope images. This preliminary classification serves as a test run for the model, allowing researchers to assess its performance and accuracy in real-world scenarios. By classifying these images, the U-Net model generates initial predictions for burned areas, tree cover, and other land cover types.
After obtaining these initial classifications, researchers conduct a thorough quality assessment of the classified images. This assessment involves comparing the model’s predictions to the actual land cover visible in the PlanetScope images. Any discrepancies or errors identified during this process are manually corrected, ensuring that the classified images accurately reflect the ground truth. These manually corrected images become valuable additions to the training dataset, further enhancing the model’s accuracy and performance.
3. Validation Assessment
After preliminary classification and quality assessment, the refined U-Net model is tested on a fixed set of validation images. This validation set consists of images that were previously and independently annotated, providing a reliable benchmark for assessing the model’s performance. By applying the U-Net to this validation set, researchers can evaluate the model’s classification accuracy comprehensively.
The process involves calculating various performance metrics, such as precision, recall, and overall accuracy, to determine the model’s effectiveness in identifying different land cover types and disturbances. These metrics offer insights into the model’s strengths and weaknesses, highlighting areas where further improvements may be necessary. The assessment results guide researchers in making informed decisions about the next steps in the active learning process.
4. Accuracy Check
Once the validation assessment is complete, researchers must decide whether the model’s classification accuracy is sufficiently high to consider the process finished. If the model demonstrates a high level of accuracy, correctly identifying and classifying land cover types and disturbances, the active learning process can be concluded. This decision is critical for ensuring that the final model meets the required standards for accurate land cover mapping.
Achieving high classification accuracy signifies that the model has effectively learned from the diverse training data and can consistently produce reliable results. However, if the accuracy falls short of expectations, further iterations of the active learning process are necessary. This iterative approach ensures continuous improvement and refinement of the model, ultimately leading to a highly accurate and reliable land cover classification tool.
5. Iteration
If the validation assessment shows that the classification accuracy is still not satisfactory, the active learning process moves into an iterative phase. In this stage, the corrected PlanetScope images from the initial classification are added to the existing training dataset. This updated dataset provides the U-Net model with new, high-quality examples to learn from, enhancing its ability to accurately classify land cover.
The U-Net model is then retrained using this expanded training set, which incorporates the latest corrections and refinements. Following this retraining, the model is used again to classify a small batch of unlabeled PlanetScope images, followed by a quality assessment and manual correction. This iterative procedure repeats until the model achieves the necessary level of classification accuracy.
Through repeated iterations, the U-Net model continually improves, becoming more skilled at identifying subtle variations in land cover and disturbances. The ultimate outcome is a highly accurate tool for large-scale land cover classification and disturbance mapping, powered by high-resolution training data and active learning techniques.
