Climate change is one of the biggest problem that humanity ever faced in their history. The causes are many and humans do have - unfortunately - a big responsibility about what is happening. Climate change is recognized as an issue that has negative efects on the ecosystem and it is mainly caused by the wild and uncontrolled deforestation of the main world forests. Another important negative efect of climate change is the sudden drying up of reservoirs and melting glaciers. Detecting deforestation and defining the causes of deforestation is an important process that could help monitor and prevent it to happen. Deforestation detection has been boosted by recent advances in geospatial technologies and applications, especially remote sensing technologies and machine learning techniques. This paper presents a land monitoring solution for deforestation, reduction of reservoirs and glaciers and the approach is based on a conceptual framework and it has been quantitatively validated on an open-source data set and qualitative evaluated on images of the Italian territory. The framework is based on machine learning and image processing techniques. It consists of three main steps, which are data preparation, model training and validation. The implementation of the proposed approach shows promising performance for detecting deforestation, reduction of reservoirs and glaciers.
lems by scientific community in the field of Deep Learning (DL), that has the objective of assigning to every pixel of an image a pre-defined given class.
In the field of image analysis, deep neural networks have been proved to be efective in solving most of the problems such as for example related to smart surveillance (e.g., recognition of people, vehicles and other moving objects) and to medical diagnostics (e.g., recognition of injuries, diseases or tumors). In all these cases, the added value of solutions based on machine learning consists in the possibility of obtaining precise information much faster and with a lower cost with respect to traditional data analysis methods. putational power and efectiveness that characterizes DLbased image analysis solutions allowed to apply them on the field of remote sensing which is the discipline that deals with the analysis and extraction of information from collected data from remote instruments such as sensors, aircraft and, in particular, satellites. In recent decades, thanks to lowering of the price for acquiring high-quality satellites images and the availability of modern state-of-the-art machine learning frameworks, it was
The feasibility of machine learning approaches in the
remote sensing domain has been demonstrated in many
applications such as for example earth observation [
With regard to deforestation and reduction of
reservoirs and glaciers, which are the general scope of this
paper, machine learning approaches can be classified into
two categories: 1) approaches detecting the location of
areas at risk of deforestation and 2) approaches
analyzing the variables that drive deforestation [
In the following next sections, we present the data set used for training the machine learning model and the architecture of the proposed solution.
The data set used for the development of the proposed
machine learning model is the LandCover.ai1 [
EPSG:2180 spatial reference system • masks with a single channel in GeoTif format with spatial reference system EPSG:2180 The data set provides a mapping of the areas represented in the images in five categories: reservoirs, forest, roads, buildings and other. The distribution of the classes within the data set is shown in Table 1. 1https://landcover.ai.linuxpolska.com
Reservoirs Forest
Roads Buildings Other
6.00 % 33.30 % 1.60 % 0.85 % 58.25 %
We processed images by dividing them into patches of size 256 × 256 with a resolution of 30 cm/pixel. Subsequently, all images that contain a total area of a target class (i.e., reservoirs, forest, roads and buildings) less than 5% were excluded. At the end of this phase, the training data set is composed of approximately 22.000 images representing all the main landscape scenarios of heterogeneous geographical areas. This data set was used for the development of territory segmentation model and it has been divided into train, validation and test for performance evaluation.
We propose a segmentation-based approach which has been proved to be very efective for this kind of tasks. Image segmentation involves converting an image into a collection of pixel regions represented by a mask or labeled image. By dividing an image into segments, you can process only the important segments of the image instead of processing it entirely. We decide to use a UNetbased network (see Figure 1) that is a cutting-edge deep (d) Segmented Area # 1
For training the model, we used a composite loss func = 2 ∗ ∑ ∑ 2
∗ + 2 + whilst the focal loss measures the degree of entropy - i.e. uncertainty - present in the classification process, scaling it by a factor capable of assigning, when estimating the value of the loss, greater weight to the more dificult examples to predict.
=1 = − ∑ ( − ) log ( ) where N represents the total number of classes in the data set. (1) (2) (d) Segmented Area # 1
The proposed solution has been evaluated by means of the following de-facto standard metrics.
IoU (Intersection over Union). It is calculated by ifrstly computing the area of overlap between the predicted and ground-truth bounding boxes (i.e., the numerator) and then by computing the area of union which is the area encompassed by both the predicted and groundtruth bounding boxes (i.e., the denominator). The IoU (see Eq. 3) is then calculated by dividing the area of overlap by the area of union that yields our final score between 0 and 1. The higher the value is, the better is the model in segmenting the image.
= ∩ ∪ (3)
F1-score. The F1-score (see Eq. 4.1) is the harmonic mean of the precision and recall. Precision refers to the number of true positives divided by the total number of positive predictions (i.e., the number of true positives plus the number of false positives). Recall, instead, refers to the number of true positives divided by the total objects of that class (i.e., the number of true positives plus the number of false negatives). The highest possible value of an F1-score is 1.0, indicating perfect precision and recall, and the lowest possible value is 0, if either precision or recall are zero.
1 - = 2 ∗ ∗ + =
2 ∗ 2 ∗ + + = =
+
+ (4)
The aim of the proposed solution was to provide a framework for mapping the territory based on specific characteristics through the use of satellite images for detecting deforestation, reduction of reservoirs and glaciers.
The presented approach is based on a the image segmentation technique which is commonly used in digital image processing and analysis. The experimentation has been conducted on a publicly available data set for a quantitative evaluation and on images of Italian territory for a qualitative evaluation. The objective is to have a tool for: • fighting against deforestation • detecting the reduction of reservoirs • detecting the reduction of glaciers In this paper, we showed the promising performance of the solution on detecting the reduction of vegetation and reduction of reservoirs. A future work will be on testing the approach on a data set having also glaciers.