Computer Vision-Based Biomass Estimation for Invasive Plants

1. Preparation of datasets

1. Selecting the research object
   1. Select test samples based on the focus of the experimental study, considering options like Mikania micrantha or other invasive plants.
2. Collecting UAV images
   1. Prepare square plastic frames of size 0.5 m*0.5 m and quantity 25-50, depending on the size of the area studied.
   2. Employ a random sampling approach to determine soil sampling locations in the study area using a sufficient number of biomass samples. Position the sample frame horizontally over the vegetation, fully encompassing the plants with a minimum separation distance of 2 m between each plant.
   3. Use a drone and a camera to form a UAV remote-sensing filming system, as shown in Figure 1.
   4. Use the UAV to plot the route within the specified study area. The route planning setup is shown in Figure 2.
      1. Establish a heading and side overlap rate of 70%, capture photos at uniform time intervals of 2 s, maintain the camera angle perpendicular to the ground at 90°, and position the camera altitude at 30 m. This resulted in continuous visible image data of the study area with a single image resolution measuring 8256 x 5504 pixels, as depicted in Figure 3.
   5. Store aerial imagery for subsequent processing with Python software for biomass estimation.
3. Collecting the aboveground biomass
   1. Collect the aboveground biomass of Mikania micrantha manually within each sample plot after completing the drone data collection. Bag them and label each bag accordingly.
      1. When collecting Mikania micrantha, prevent sample plots from moving. First, cut the Mikania micrantha along the inside edge of the sample plot.
      2. Then, cut the rhizome of the Mikania micrantha from the bottom. Remove any dirt, rocks, or other plants that are mixed in. Finally, bag and label the samples.
   2. Bring the collected samples of invasive plants from step 1.3.1 to the laboratory. Air-dry all the collected samples to evaporate most of the moisture.
   3. To further remove moisture from air-dried samples, use an oven. Set the temperature to 55 °C. Dry the samples for 72 h, then weigh each sample on an electronic balance and record the biomass data in grams (g).
      1. Place the electronic balance in an undisturbed environment, weigh, calibrate, and continue weighing. Place bags of Mikania micrantha on the electronic balance, wait for the readings to stabilize, and record the readings.
      2. Weigh the Mikania micrantha every hour until the mass no longer changes and record the reading minus the weight of the bag as the measured mass of that sample. Calculate the aboveground biomass of the invasive plant using the formula below:
         
         where B represents the biomass of Mikania micrantha in grams per square meter (g/m²), M is the weight of the measured Mikania micrantha, measured in grams (g), S corresponds to the area of the sample plot in square meters (m²).
4. Creating a dataset
   1. Extract the RGB image corresponding to the sample image from the original UAV image. Divide it into a grid of 280 × 280 pixels using Python programming (Supplementray Figure 1).
   2. Segment the raw image data into smaller images of the same size as the sample images using Python programming. Use the sliding window method for segmentation, setting the horizontal and vertical steps to 280 pixels.
   3. From the small images segmented in step 1.4.2, randomly select 880 invasive plant images and 1500 background images to create a dataset. Then, split this dataset into training, validation, and test sets in a 6:2:2 ratio (Supplementray Figure 2).

2. Identification of Mikania micrantha

1. Preparing the software
   1. Go to Anaconda's official website (https://www.anaconda.com/) and download and install Anaconda. Then, head to PyCharm's website (https://www.jetbrains.com/pycharm/) and download the PyCharm IDE.
2. Creating a Conda environment.
   1. Open the Anaconda Prompt command line after installing Anaconda, then type conda create -n pytorch python==3.8 to create a new Conda environment. After the environment is created, enter conda info –envs to confirm that the pytorch environment exists.
   2. Open the Anaconda Prompt and activate the pytorch environment by entering conda activate pytorch. Check the current (Compute Unified Device Architecture)CUDA version by typing nvidia-smi. Then, install PyTorch version 1.8.1 by running the command conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=11.0 -c pytorch.
3. Runs for model recognition
   NOTE: Use PyTorch to build the Mikania micrantha recognition model used in this paper. The network model employed is ResNet101²¹, which remains consistent with the original paper in its architecture. Modifications are made to the network's output section to meet the requirements for chamomile recognition.
   1. Preprocess the images to prepare them for model input. Resize the images from 280 x 280 pixels to 224 x 224 pixels and normalize them to ensure they meet the model's size requirements using the following code:
      transform = transforms.Compose([
      transforms.Resize((224, 224)),
      transforms.ToTensor(),
      ​transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
   2. Perform image feature extraction and reduce dimensionality using a convolutional neural network.
      1. First, initialize the convolutional layer for the initial feature extraction through self.conv1. With this convolutional layer, the original image is convolved into a feature map with self.in_channel channels for extracting initial features (Supplementary Figure 3A).
         NOTE: Advanced features are extracted in a convolution operation on the residual passes. These layers are produced by invoking the _make_layer function, which comprises a sequence of residual blocks. Each residual block consists of convolution, batch normalization, and activation functions to gradually extract sophisticated features (Supplementary Figure 3B).
      2. Use the layer's function to modify the channel number for dimensionality reduction via 1×1 convolution. This operation decreases the computational load while preserving significant features (Supplementary Figure 3C).
         NOTE: Overall, ResNet101 performs feature extraction by using various convolutional layers, and dimensionality reduction is achieved through 1×1 convolutional layers within the residual block. This approach allows the network to learn features more deeply and avoid the issue of gradient vanishing, thus enabling more efficient learning and representation of image features for complex tasks.
      3. After convolutions and pooling operations, input the high-quality features into a fully connected layer.
         NOTE: In the ResNet architecture, feature extraction takes place in the convolutional layer. These features are subsequently sent to the fully-connected (FC) layer for classification (Supplementary Figure 4). The operation self.avgpool(x) performs adaptive average pooling to reshape the tensor to a fixed size. The operation torch.flatten(x, 1) spreads the tensor into a one-dimensional vector, and self.fc(x) applies the fully connected layer to the flattened vector, ultimately serving as the final step for classification. This process effectively passes the extracted features through the convolutional layer, transforming them into a format suitable for classification via the fully connected layer.
      4. Use the Softmax function to obtain the final output based on the three classification requirements.
   3. Train a multi-class recognition model with the dataset from step 1.4. Set the number of iterations to 200 and an initial learning rate of 0.0001. Reduce the learning rate by a third every 10 iterations with a batch size of 64. Save the optimal model parameters automatically after each iteration (Supplementary Figure 5).
   4. Employ a meticulously trained recognition model and systematically traverse the original image from step 1.2.2 for identification purposes.
      1. Configure horizontal and vertical steps precisely at 280 pixels, resulting in the generation of a comprehensive distribution map highlighting the presence of invasive flora within the boundaries of the study area. Present the selected results visually as shown in Figure 4.
         NOTE: The initial image is preprocessed by segmenting it into smaller chunks, classifying each chunk using a trained deep learning model, and combining the results into an output image. If a chunk is classified as an invasive plant, the corresponding location on the output image is set to 255. The resulting output image is saved as a grayscale image file. The specific implementation code is shown in Supplementary Figure 6.

3. Estimation of invasive plant biomass

1. Perform simple data augmentation with the RandomResizedCrop and RandomHorizontalFlip functions (Supplementary Figure 7) to extend the image set created in step 1.2 and extract the six vegetation indices commonly used for estimating biomass, which are RBRI, GBRI, GRRI, RGRI, NGBDI, and NGRDI. Refer to Table 1 for the calculation formulas for these indices.
2. Create a K-nearest neighbor regression (KNNR)²² model using the output of the model to ensure precise estimation of the biomass of invasive plants. Use the extracted vegetation indices as inputs for the estimation model.
3. Use the coefficient of determination R² and Root Mean Square Error (RMSE)²³ to assess the model's accuracy, which is calculated as follows:
   
   
   NOTE: The K-Nearest Neighbor Regression (KNNR) algorithm is a nonparametric machine learning technique used to solve regression problems. Its fundamental concept is to predict outcomes by determining the closest K neighbors in feature space based on input sample distances. Key benefits of using KNNR include its simplicity and ease of understanding, and it requires no training phase. Additionally, KNNR does not make excessive assumptions about the data's distribution. KNNR can be applied in regression issues to anticipate continuous objective variables and precisely evaluate the biomass of invasive plants.
4. Employ the aboveground biomass estimation model chosen in step 3.2 and scan through the invasive plant distribution map from step 2.3.4 with horizontal and vertical strides of 280 pixels.