IDTReeS: Integrating Data Science with Trees and Remote Sensing

By Rahul Koli, Computer Engineering, VESIT

Introduction

Understanding and managing forests is a crucial task to understand the potentially mitigating the effects of climate change, invasive species, and shifting land use on natural systems and human society. However, collecting data on individual trees in the field is expensive and time consuming, which limits the scales at which this crucial data is collected. This data science project focuses on inferring ecological information from remote sensing data.

Dataset

The data for this evaluation is geospatial data, which is data that is connected to a physical location on Earth’s surface. The remote sensing data are generated by the NEON Airborne Observation Platform (AOP). These data are distributed as raster and ‘point cloud’ vector formats. Remote sensing data comes from either passive or active systems.

This project includes data from both systems.

Passive system: record energy from the sun that is reflected from Earth’s surface. The sensors can measure the amount of energy reflected at different wavelengths. Passive remote sensing data includes RGB and hyperspectral datasets.

Active system: Record energy that is reflected from Earth’s surface from the system itself. The system emits pulses of energy towards Earth’s surface and measures time and intensity of energy that returns to the sensor. For this competition, active remote sensing data includes light detection and ranging (LIDAR).

Figure 1: Dataset used to train the model

Methodology

In this project we used deep forest model. Current deep learning models are mostly built upon neural networks, i.e., multiple layers of parameterized differentiable nonlinear modules that can be trained by backpropagation. In this project, we explore the possibility of building deep models based on non-differentiable modules. The DeepForest model uses the keras—retinanet evaluate method to score the images. It also needs the “annotations.csv” file in the format specified. The evaluation file can be run as a call back during training by setting the config file. It will allow the user to see evaluation performance at the end of each epoch. DeepForest uses the Keras for saving workflow, where users can save the entire model architecture, or just the weights. We have done this task in two parts delineation and classification.

1. Delineation to locate individual trees crowns (the top, sun-exposed portion of the tree visible from above) in remote sensing data.

2. Classification to determine the taxonomic species identity (i.e., the type/category) of each individual tree in remotely sensed data.

Figure 2: The format of annotations.csv file

The delineation of tree crowns was successfully detected as shown in the images below.

Figure 4: Tree crown detected from the MLBS region 
Figure 5: Tree crown detected from the MLBS region 

Conclusion

Finally, we achieved the task that is to identify the top of the tree crowns in remote sensing data to their taxonomic species. In addition to its utility for the domain, this task represents a challenging version of general classification problems because it involves classifying different species with very similar spectral signatures and categorizing data.

YouTube: IDTREES

Blog: IDTReeS

TEAM:

Marcia Rajang
Rahul Koli
Kavya R Nair
Pooja Rahate
Umadevi Kollabatthina

MENTOR: Dr. Shakti Sharma

Leave a Reply

Your email address will not be published. Required fields are marked *