TREE DETECTION AND CLASSIFICATION THROUGH AERIAL IMAGERY USING DEEP NEURAL NETWORKS

Project Title

TREE DETECTION AND CLASSIFICATION THROUGH AERIAL IMAGERY USING DEEP NEURAL NETWORKS

Project Area of Specialization Artificial IntelligenceProject Summary

Automatic detection and classification of trees by using remotely analyzed information have been a dream of the many scientists, and land use administrators. The motivation for this problem comes from pollen tree excavation issue, automated 3D town modeling, urban planning, and forestation, within which such information is employed to come up with the models. Here, we offer an automatic methodology for individual tree detection and classification through aerial imagery using unmanned aerial vehicles (UAV), which is a rapidly evolving, cost-effective and economical technology. Firstly, the model is trained for the purpose of tree detection per image pixel by assigning a {tree, non-tree} label to each pixel in an aerial image. Afterward, the output is refined into clean segmented image based upon which, we implement pattern matching to locate the separable tree crowns, which are then classified on the basis of tree species type with our algorithm. We have verified the algorithm on many gathered aerial pictures across varied zones of a district and have confirmed excellent quality results with good scalability of our proposed methodology. In contrast, most of the formerly done work used costly hardware like multispectral images for tree detection and classification. Thus, our proposed technique has the potential to classify individual trees in an exceedingly cost-effective manner. This will be a useful tool for several forest researchers, management, and also for the concerned government bodies to detect and excavate pollen trees, to fight with this seasonal pollen allergy war.

Project Objectives

• To attain an automated methodology for the detection and classification of trees in a region from aerial imagery obtained by UAV (unmanned aerial vehicle)
• To design and formulate a neural network framework with the help of structural algorithm of machine learning and implication of pixel-level classifier and utilize it for objects that have high entropy, for instance, trees, grass, river, etc. that are natural objects.
• To attain an automated methodology for the detection and classification of trees in a region from aerial imagery obtained by UAV (unmanned aerial vehicle)
• To design and formulate a neural network framework with the help of structural algorithm of machine learning and implication of pixel-level classifier and utilize it for objects that have high entropy, for instance, trees, grass, river, etc. that are natural objects.
• To boost the work of forestation and meteorological departments.

Project Implementation Method

Data Set Acquisition

For the purpose of data set formation, we have covered aerial imagery with the help of our drone of various areas of Rawalpindi and Islamabad.

Feature Extraction and Detection:

 First of all, we have gathered the aerial imagery of various areas to generate our Dataset which is then processed for the formation of ortho maps.

Now, these images will be passed through our designed sequential model based on deep neural network for tree detection purpose.

The features that we are considering are as follows:

• Red color proportion
• Texture using Gabor filter
• Intensity value
• Saturation value
• Hue concentration
• Green color proportion
• Blue color proportion

These features will be extracted per pixel with the help of our algorithm and a vectorized feature vector matrix will be generated on the basis of the features selected.

Afterward, we will generate binary images with the help of our produced ground truths.

Data Vector is made by the combination of feature and label vectors, which will be passed for model training.

The trained model then tested with the defined and fixed manner of data streams of the aerial imagery for getting the mark of tree and non-tree on the pixels of the image as output. The output of the model is the prediction which is carried on the same model trained parameters. Thus the tree detection of our proposed methodology has been completed.

Classification

The output generated from the previous part is then refined and vigorously multiplied with the original dataset image of the same tile for getting the segmented imagery dataset, it is then sliced down into respective tiles for its classification process. The slicing of the imagery dataset is done in a defined order to maintain the same pixel’s ration and its picture definition undisturbed.

Now, the labeling process is carried out on each and every tile (sliced images) to form an Excel file which is then converted to array form afterward of the labeling. The imagery dataset is also maintained in a huge encoded vector in categorical order already for the classification purpose. After, the model description on the defined platform, the CNN model is trained over the training dataset of the developed maps (which are previously optimized according to the requirements of the designed architecture). The trained model is then tested and validated by exposing over the same classes of the dataset and we get the results of the classification from the predicted output. The output is contained of the encoded vector which contained the prediction score of the categories which is then taken as maximum argument form for each category prediction. It is then relatively compared along with another prediction score of the other classes to unified a single output of the trained model.

Benefits of the Project

•Female Pollen Tree Excavation

•Generalized Report Generation

•Enhancement of  the possible Map versions

•Utilization for Forestry and Forest Management

•Further advancements along with minor improvements to analyze Agricultural Areas

Technical Details of Final Deliverable

The designed deep learning model for detection has the architecture with the following characteristics:

• Model = Sequential (Keras library implementation)
• Evaluation is also performed on the same measuring metrics
• Batch size of 100000
• Metrics= Accuracy
• Optimizer = Adam
• Loss = binary cross-entropy
• The model compilation contains loss, optimizer and measuring metrics
• While the last layer contains “Sigmoid” function
• The initial layering contains of “Relu” function
• 3rd layer only contains one neuron
• 2nd layer with 50 hidden units
• 1st layer with 100 hidden units
• 3 layered model

The implied and deployed architecture of the convolutional neural network (CNN) for classification has the following characteristics:

• Model = Convolutional Sequential Model (Keras library)
• Optimizer = Adam
• Epoch = 10
• Loss = Categorical cross-entropy
• The model compilation contains loss, optimizer, and measuring metrics
• While the last layer contains “Softmax” function
• The initial layering contains “Relu” function
• The dense layer contains neurons of equal labels of classes
• Flattening layer
• 3rd layer with further downed size input from previous output
• 2nd Max pooling2D with pool-size
• 2nd layer with input size of previous output size
• 1st Max pooling2D with pool-size
• 1st layer with an input size of 256 x 256
• Pool-size = 2 (Only for Max pooling)
• Dropout = 0.2
• Filters = 40
• Kernel size = 3 x 3
• Batch size of 100
• Input shape of 256 x 256 x 1 (greyscaled)
• Image reshaping and resizing is performed for data fitting according to model
• Convolutional 2D 3-layered model with 1 dense layer
• Metrics= Accuracy

Final Deliverable of the Project Software SystemType of Industry Agriculture Technologies Artificial Intelligence(AI)Sustainable Development Goals Climate ActionRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	68700
Intel movidius neural compute stick	Equipment	2	20000	40000
High speed processing server	Equipment	1	28700	28700