Introduction to Deep Learning

Overview

Goal: Provide an introduction to deep learning concepts and how they can be applied to enhanced forest inventories (EFIs)

Overview

By the end of this section, you should be able to:

  • Explain the role of AI in EFI modelling
  • Distinguish AI, machine learning, and deep learning
  • Describe the structure of a deep neural network
  • Explain how models learn (loss, gradients, epochs)
  • Compare image-based vs. point-based deep learning
  • Use core deep learning terminology correctly
  • Recognize common forestry-specific challenges

What Is Deep Learning?

Artificial Intelligence → Machine Learning → Deep Learning

  • Artificial Intelligence (AI): Systems that mimic human decision-making
  • Machine Learning (ML): Models that improve through data-driven experience
  • Deep Learning (DL): ML using neural networks with many hidden layers

Deep Neural Networks: Key Concepts

Neural Network Components

  • Layers: Input → hidden → output
  • Depth: Number of layers
  • Width: Number of neurons per layer

Neural Network Components

Trade-offs:

  • Increases in depth/width → increase representational power
  • Increases in depth/width → increase computation, overfitting risk

The Neuron

  • Fundamental unit of a neural network
  • Computes a weighted sum of inputs plus a bias

\[ \Large y = w_1x_1 + w_2x_2 + b \]

  • Weights & biases are learned during training

Activation Functions

  • Applied to neuron output
  • Introduce non-linearity
  • Enable learning of complex relationships

Note

There are many types of activation functions, and you are encouraged to research which one would be best for your specific use case.

How Do Neural Networks Learn?

Prediction–Feedback Loop

The prediction-feedback loop is what makes learning possible as a deep neural network adapts to the patterns within the data and generalizes from them.

Goal: Loss, Gradients, and Epochs

Learning is driven by the loss function which guides the adjustments of the weights/biases over a number of epochs.

  • Loss function: Measures prediction error
  • Gradients: Sensitivity of loss to small changes in weights/bias
  • Epoch: One full pass through the training data

Common Deep Learning Tasks

Learning is guided by defining what the task to be accomplished is.

  • Classification
  • Regression
  • Segmentation

Task: Classification

Goal: Assign inputs to discrete classes

Forestry examples:

  • Tree species mapping
  • Land cover classification
  • Disturbance detection

Task: Regression

Goal: Predict continuous values

Forestry examples:

  • Above-ground biomass
  • Canopy cover / LAI
  • Tree height, crown size

Task: Segmentation

Goal: Assign labels to every pixel or point

Forestry examples:

  • Tree crown delineation
  • Stand boundaries
  • Structural mapping

Data: Training, Validation, Testing

Learning happens through patterns identified in the data.

  • Training: Learn parameters
  • Validation: Tune hyperparameters, prevent overfitting
  • Testing: Final, unbiased performance estimate

Note

Never mix training, validation, and testing data. Mixing or reusing data between these datasets can lead to misleading performance metrics and poor generalization.

Image-Based vs. Point-Based Deep Learning

Two dominant approaches in remote sensing:

  • Image-based methods
  • Point-based methods

Image-Based Deep Learning

  • Operates on raster imagery
  • Learns spatial patterns and textures
  • Common inputs: RGB, multispectral, hyperspectral

Convolutional Neural Networks (CNNs)

  • Core image-based architecture
  • Use sliding filters (kernels)
  • Detect edges → textures → objects

Point-Based Deep Learning

  • Operates directly on 3D point clouds
  • No regular grid structure
  • Learns from spatial relationships

Set Abstraction

  • Common point-based architecture
  • Group nearby points instead of grids
  • Learn local geometry → shapes → objects

Core Terminology

Establishing shared language before workflows and modeling.

Parameters vs. Hyperparameters

  • Parameters: Learned values (weights, biases)
  • Hyperparameters: User-defined settings
    • Learning rate
    • Batch size
    • Epochs
    • Model depth

Inputs, Labels, Learned Features

  • Input features: What the model sees
  • Labels: What the model predicts
  • Learned features: Internal representations
  • Logits: Outputs of the final layer of a model

Data Augmentation

  • Artificially increases training diversity
  • Improves robustness and generalization
  • Especially useful with limited labels

Batches and Batch Size

  • Data processed in small subsets (batches)
  • Batch size: Samples per training step

Architecture vs. Weights

  • Architecture: Model/network structure (fixed)
    • Number and type of layers
    • How neurons/layers connect
  • Weights: Learned numerical values (changes)
    • Updated during training
    • Learned based on data

Training Loop

Repeated cycle until optimal point reached:

  1. Forward pass (predict)
  2. Loss computation (assess)
  3. Backward pass (feedback)
  4. Parameter update (adjust)

Transfer Learning and Fine-Tuning

  • Transfer learning: Start from pretrained model
  • Fine-tuning: Adapt weights to your domain

Common in forestry due to limited labeled data.

Deep Learning in Forestry: Challenges

Forestry data introduce domain-specific constraints.

Key Challenges and Considerations

Challenge Impact Common Solutions
Sensor & site variability Poor generalization Fine-tuning, domain adaptation
Spatial autocorrelation Inflated accuracy Spatial CV, sampling strategies
Class imbalance Bias toward majority classes Class weighting, loss design
Synthetic benchmarks Limited real-world transfer Domain adaptation
Limited labels Reduced robustness Augmentation, transfer learning

Important

Human expertise is essential in AI-derived forest inventories because models do not define objectives, meaning, or validity on their own. Forestry knowledge is required to select appropriate data, define training and modelling approaches, and interpret outputs/results ensuring they are scientifically sound, enabling deep learning to serve as a powerful tool for informed forest management decision-making.