FR3116925A1 - Method for detecting photovoltaic installations on an image by deep learning - Google Patents

Abstract

METHOD FOR DETECTING PHOTOVOLTAIC INSTALLATIONS ON AN IMAGE BY DEEP LEARNING Method (500) for detecting objects (10) on a previously acquired aerial or satellite image (100), implementing a deep learning model (200) executed on a computer and comprising: a step (520) of cutting image into a plurality of thumbnails adapted to an input format of the model; a step (530) of detecting objects (10) on the image comprising an activation calculation (531) on each thumbnail and giving a classification (532) and a CAM class activation map (533) of said thumbnail; a step (540) of grouping adjacent thumbnails of the same classification into blocks; a step of filtering (550) the blocks; and a step (560) of estimating the geographical coordinates and the surface of each object detected from a segmentation of said blocks. The detected objects are for example photovoltaic installations. Figure for the abstract: figure 6

Description

Method for detecting photovoltaic installations on an image by deep learning

The present invention belongs to the field of image processing by artificial intelligence and techniques commonly known as computer vision, in particular the detection of objects by automatic learning ( Machine Learning ) or deep learning ( Deep Learning ). The invention relates more particularly to a method for detecting photovoltaic installations on an aerial or satellite image by deep learning, implementing a network of artificial neurons.

The invention finds a direct application in the renewable energy sector to map the solar park of a given geographical region.

State of the art

The deployment of solar photovoltaic energy is accelerating worldwide due to the rapid reduction in costs and significant environmental benefits compared to generating electricity from fossil fuels. Due to the decentralized nature of photovoltaic installations, data concerning them remains difficult to access for the entire solar park in a given region. For example, information on the location and power of each installation within a fleet is often incomplete, or even unavailable, including in the most industrialized countries oriented towards renewable energies.

In France, the existing data is generally in the form of agglomerated lists at the level of municipalities and, in the majority of cases, does not specify the size and location of each installation.

There is therefore a real need to have an accurate and complete database gathering the information necessary for the proper monitoring of the evolution of a solar park in a given region. Such a database would prove extremely useful, for example, to monitor the development of a solar park in each category of installation, from photovoltaic power plants to individual panels for individuals, and to estimate its energy production.

Currently, remote sensing makes it possible to map photovoltaic installations in a region, but with the intervention of human operators to analyze the images. Consequently, human tasks can quickly become tedious, even insurmountable, when the regions concerned present large areas.

It is known that artificial intelligence combined with imagery, and more particularly satellite or aerial imagery, makes it possible to automate object recognition and detection operations quickly and accurately.

CN111178316A, as an example, describes a general method for classifying land cover from high-resolution remote sensing images, based on deep learning. The method is used for the automatic search of a suitable convolutional neural network architecture for a specific data set.

The teaching of this document nevertheless remains theoretical and does not offer any possibility of application to the detection of photovoltaic installations.

Document CN111191500A concerns a method for identifying photovoltaic roofs based on image segmentation by deep learning. The method consists of combining several segmentation models to classify different types of roofs. The objective of this method is to automate certain tasks carried out manually before and thus save considerable time.

Specifically, this method first obtains remote sensing imagery from a mapping service provider and manually annotates the data to build an image segmentation dataset, trains the U-Net network on three roof textures (cement , colored steel tiles and photovoltaic panels), discovers the typical prediction errors of the model, and performs continuous iterative training of the model. After each type of training is completed, the method determines the best classification threshold for each set of tests. Then, it uses Conditional Random Fields ( CRFs ) to optimize the model results based on the prediction results of the three models in the final combination. Finally, the optimized model is used to segment the remote sensing images and recognize photovoltaic panels in the specified area.

However, the use of CRFs for model optimization has a major drawback which is their lack of robustness since they have been tested only on well-structured data.

DeepSolar, whose reference is given in the description, is another convolutional neural network, developed by researchers at Stanford University, with the aim of detecting photovoltaic panels on high resolution images (less than 30cm). This model takes as input images cut into pieces ( Tiles in Anglo-Saxon terminology) and classifies each piece according to whether it contains photovoltaic panels or not.

The present invention implements the DeepSolar model and provides significant improvements to basic processing. The additional details of how DeepSolar works will therefore be given directly in the description below.

It should be noted, however, that the DeepSolar model, as designed, is used to create an agglomerated dataset over a limited area so that the total area of the photovoltaic panels in said area can be compared against certain meteorological, economic and other parameters.

Presentation of the invention

The present invention aims to overcome the drawbacks of the prior art set out above and proposes a solution making it possible to create an updatable data set grouping together the photovoltaic installations of a given region thanks to the DeepSolar model, or a suitably trained similar model, and different image processing techniques.

To this end, the subject of the present invention is a method for detecting objects on a previously acquired aerial or satellite image, implementing a deep learning model executed on a computer and comprising: a step of cutting the image into a plurality of thumbnails adapted to an input format of the model; a step for detecting objects in the image comprising an activation calculation on each thumbnail and giving a classification and a CAM class activation map of said thumbnail. Advantageously, this method further comprises: a step of grouping adjacent thumbnails of the same classification into blocks; a block filtering step; and a step of estimating the geographic coordinates and the surface of each object detected from a segmentation of said blocks.

More particularly, the method is intended for high resolution images, preferably with a resolution of less than 30cm.

According to an advantageous embodiment, the classification is Boolean and results for each thumbnail from the comparison of the activation of said thumbnail with a threshold chosen by an operator.

Preferably, the chosen threshold is that maximizing a function whose variables are the precision and recall metrics calculated on a set of classification calibration data.

More preferably, the threshold chosen maximizes a function also taking as a variable a recall metric per object, defined as the percentage of objects having at least one part detected on a thumbnail of a calibration image among the objects visible on said image.

According to a particularly advantageous embodiment, the step of filtering the blocks consists in filtering the blocks of the visible thumbnails and/or the blocks of the associated CAM activation cards according to at least one criterion among: the size, the color and the texture of the detected object, to eliminate false positives.

According to one embodiment, the segmentation comprises k-means partitioning of the visible thumbnail blocks and pixel analysis of the associated CAM activation maps.

According to the invention, the detected objects are for example solar and/or photovoltaic panels which can be grouped into installations.

According to this example of application, the method further comprises a step of determining a nominal power for each photovoltaic installation detected.

The invention also relates to a computer program product characterized in that it comprises a set of program code instructions which, when they are executed by a processor, implement a detection method as presented .

The fundamental concepts of the invention having just been explained above in their most elementary form, other details and characteristics will emerge more clearly on reading the description which follows and with regard to the appended drawings, giving by way of illustration non-limiting example an embodiment of a method for detecting photovoltaic installations on images by deep learning, in accordance with the principles of the invention.

Presentation of drawings

The figures are given for purely illustrative purposes for the understanding of the invention and do not limit the scope thereof. The different elements are represented schematically and are not necessarily to the same scale. In all the figures, identical or equivalent elements bear the same reference numeral.

It is thus illustrated in:

: an aerial image of a geographical area in which photovoltaic installations are visible;

: a cutout of the image of the in thumbnails;

: examples of thumbnails cut out of the image;

: a functional diagram of the deep learning model used in the detection method according to the invention;

: a block diagram of the deep learning model, with its inputs and outputs;

: the main steps of the detection method according to one embodiment of the invention;

: an example of colorimetric correction of a real image;

: a division into thumbnails of the corrected image of the ;

: the classifications and the class activation cards obtained for the tiles of the during the detection step;

: a manually annotated image portion, representing classification threshold calibration data;

: the evolution of different statistical metrics for different values of the classification threshold;

: a grouping of the positive thumbnails into blocks according to the invention;

: a comparison between a false-positive and a true-positive according to a texture criterion;

: a binary activation of a positive block from the segmentation of the visible block and the associated activation map;

: a graphical distribution of photovoltaic power plants according to their surface and their power;

: the construction of a piecewise linear transfer function from the graph of the .

Detailed description of embodiments

It should be noted that certain technical elements well known to those skilled in the art are described here to avoid any insufficiency or ambiguity in the understanding of the present invention.

In the embodiment described below, reference is made to a method for detecting photovoltaic installations on aerial or satellite images by deep learning, intended primarily to establish a precise mapping of photovoltaic installations in a given geographical area. This non-limiting example is given for a better understanding of the invention and does not exclude the adaptation of the method to the detection of other recognizable objects on aerial views such as swimming pools, sports fields and others, especially when no database exists for the objects considered.

In the rest of the description, the terms “network” and “model” designate by extension respectively an artificial neural network and a deep learning model, and each refers to DeepSolar unless otherwise indicated. Furthermore, the terminology used should in no way be interpreted in a limiting or restrictive manner, but simply in conjunction with particular embodiments of the invention.

The method makes it possible to detect photovoltaic installations on aerial images by executing an existing network which takes as input aerial images in the visible spectrum, captured by means of a suitable acquisition system, airborne or on board a Earth observation satellite, and outputs a cartographic file containing the geographical coordinates, the surface and an estimate of the nominal power, called peak power, of each photovoltaic installation detected.

There represents an aerial image 100 of a given geographical area, on which photovoltaic panels 10 of different shapes and sizes are visible. Such an image may come from a source such as an online mapping service.

Preferably, the image 100 is of high resolution, for example of resolution lower than 30 cm, to be adapted to the deep learning model used.

Indeed, the detection method, according to the embodiment described, implements the DeepSolar model available in open source , initially designed to map the photovoltaic panels deployed on American territory and thus constitute a reliable and scalable database. This model is based on the detection of photovoltaic panels on a high-resolution aerial image using a method detailed in the article: Jiafan Yu, Zhecheng Wang, Arun Majumdar , Ram Rajagopal , " DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States”, Joule, Volume 2, Issue 12, 2018, Pages 2605-2617, ISSN 2542-4351 .

The Deep Solar model takes as input high-resolution images of an area covering an area of approximately 480m ² , said model having been trained on images of this size. Consequently, the initial images 100, which are larger, must be cut beforehand into pieces, which will be called thumbnails, compatible with the input format of the model.

There represents the image 100 through a cutting grid making it possible to obtain thumbnails T of the same size. Preferably, each thumbnail T has a square shape with sides of 22m to match the input format of the model.

There shows a detail view of three isolated thumbnails T1, T2 and T3 of image 100 according to the division of the . In this example, only thumbnails T1 and T3 include parts of photovoltaic panels.

Before describing the main steps of the method according to the invention, it is recalled that the DeepSolar network is of the CNN ( Convolutional Neural Network ) convolution type and uses the images 100 according to the diagram briefly described below.

There represents the main operations performed by the model on the input thumbnails T1, T2 and T3 from frame 100. First, the input format thumbnails are imported into the model. A specific classifier is then applied to the thumbnails to determine the positive thumbnails (containing at least part of a photovoltaic panel) and the negative thumbnails (containing no part of a photovoltaic panel). The results of the classification are used to identify the positive vignettes, here T1 and T3, following which a segmentation by semi-supervised learning is applied, via specific layers of the network, to said positive vignettes. This makes it possible to generate a class activation map (CAM) for each positive thumbnail, in shades of gray where the lighter pixels indicate a greater probability of the presence of a part of the photovoltaic panel. Finally, a binary activation BA is obtained for each CAM by applying a threshold (value of the activation on the panel pixels). Thus, the model makes it possible to obtain both the size of the panels and the number of photovoltaic installations.

There represents the overall implementation of the detection method according to the invention, in which an aerial image 100 is divided into n*m thumbnails Tij, with i an integer between 1 and n, and j an integer between 1 and m, n and m being two integers greater than 1. The thumbnails form a matrix given as input to the model 200. For each thumbnail Tij, the model gives an output Sij comprising the GEO geolocation, the CLASS classification and the CAM class activation map associated with said thumbnail.

These outputs then make it possible to know the area and the peak power of each photovoltaic installation detected.

There represents a method 500 for detecting photovoltaic installations on an aerial image by deep learning comprising:

• an optional step 510 of preprocessing a previously acquired image, consisting of a colorimetric correction for example, to obtain a corrected image adapted to the deep learning model used, in this case DeepSolar;
• a step 520 of cutting the corrected image into thumbnails of a size adapted to the input format of the model;
• a step 530 of detecting photovoltaic (PV) panels by executing the deep learning model to identify the positive thumbnails and generate the associated CAMs;
• a step 540 of grouping neighboring positive thumbnails, and associated CAMs, into blocks;
• a step 550 of filtering the blocks according to different criteria;
• a step 560 of estimating the geographical coordinates and the surface of each detected PV installation;
• a step 570 of determining the peak power of each detected PV installation; And
• a step 580 for eliminating any remaining false positives.

The preprocessing step 510 consists in correcting the colorimetry of the initial image so as to obtain a corrected image corresponding better, in terms of contrast and brightness, to the type of image on which the model was trained.

In the case of the DeepSolar model, the training images are aerial images of the American territory, coming from the application programming interface of the Google Maps online service. Thus, the more the images provided to the model have characteristics (resolution and colorimetry) close to the images used to train the said model, the higher the detection accuracy of the photovoltaic installations on these images will be.

There represents an initial image 100-0 before correction and a corrected image 100 obtained after the preprocessing step 510.

The contrast and brightness correction can be performed by a specific algorithm or by a suitable method already implemented in the programming language used. For example, with the Python language, the correction step can be performed with the rescale_intensity method of the exposure module of the free image processing library Scikit-image.

Of course, the preprocessing step 510 is not necessary when the initial image has a colorimetry close to that of the reference images on which the deep learning model was trained, in which case the corrected image is directly initial picture.

The cutting step 520 then makes it possible to put the corrected image in the input format of the DeepSolar model. Thus, the image is cut into geolocated thumbnails of the same size as explained above. A cutting grid is obtained in the form of a matrix, on which the thumbnails are located by their positions in the matrix, positions which can be indicated by numerical indices of rows and columns.

There represents the image 100 cut into thumbnails T of the same size, the cutting grid being represented by continuous white lines. Preferably, the thumbnails T obtained have a square shape with side a.

The size of the T thumbnails should be consistent with the size of the DeepSolar model training images. In this case, this model was trained with square images of 22m side. We will therefore preferably choose a = 22m.

In addition, during the cutting step 520, a cartography of the geographical area covered by the initial image can be used to remove the "unnecessary" thumbnails, which completely correspond to surfaces on which the installation of an installation photovoltaic is impossible such as the ocean.

Once the input tiles are defined, these are injected into the DeepSolar model for the detection of photovoltaic installations.

The step 530 for detecting photovoltaic panels on the input thumbnails consists of executing the DeepSolar model which provides for each of said thumbnails a classification and a CAM class activation map.

Indeed, the DeepSolar model calculates during a sub-step 531 a global activation on each thumbnail. This activation corresponds to a probability of presence of at least part of a photovoltaic panel on the thumbnail, or to a confidence index on said presence. Then, the model subsequently operates a classification and a generation of CAM of the thumbnail considered during the sub-steps 532 and 533 represented on the .

Classification 532 consists of assigning a Boolean value (positive or negative) to each thumbnail according to the presence of at least one part of a photovoltaic panel on said thumbnail: positive = at least one part of a photovoltaic panel is visible on the thumbnail; negative = no part of the panel is visible. Indeed, the classification of a thumbnail is determined by the comparison of the global activation calculated in step 531 with a threshold 430 set by the operator, such that a global activation greater than the threshold implies a positive value of the classification. In other words, when the probability of the presence of at least part of a photovoltaic panel on the sticker exceeds a certain value set by the operator, the model considers that the sticker is positive and classifies it accordingly.

The generation of the CAMs 533 consists in exploiting the activation calculation having led to the classification of the thumbnails in order to create a CAM for each thumbnail. The CAM corresponds to a grayscale image representing the activation of each pixel of the thumbnail considered, the activation being an integer between 0 and 255.

There represents the classifications (0 or 1) and the CAMs obtained after the detection step 530 for the thumbnails of the image 100 of the .

The choice of the threshold 430 is therefore crucial for the step 530 because it determines the accuracy of the detection carried out.

The threshold can be set using a set of test data during steps that may be similar to a calibration of the model. The test data must be varied enough to be able to determine a threshold making it possible to improve the precision of the detection at step 530.

To do this, a portion of the corrected image, including photovoltaic installations of different sizes, is chosen. The thumbnails associated with this portion of the image are, on the one hand, manually annotated to indicate their classifications (presence or absence of a panel) and, on the other hand, provided as input to the DeepSolar model.

There represents an image portion comprising approximately 25,000 manually annotated thumbnails. The blackened areas correspond to thumbnails on which photovoltaic panels are visible. Thus, these data constitute calibration data, called truth data, which will make it possible to determine the optimal threshold.

From this set of truth data, three metrics are then calculated for different threshold values: precision, recall, and recall per PV plant. These metrics are defined as follows:

• Accuracy corresponds to the percentage of “true” installations among the installations detected. In other words, the number of true-positives compared to the sum of true-positives and false-positives;
• The recall ( Recall in Anglo-Saxon terminology) corresponds to the percentage of installations detected among the existing installations. In other words, the number of true-positives compared to the sum of true-positives and false-negatives;
• the recall per installation corresponds to the percentage of installations having at least one panel detected on a thumbnail of the image among the existing installations.

The recall by installation makes it possible to adjust the choice of the threshold more finely because it is less discriminating than the classic recall.

Thus, the value retained for the threshold is that which makes it possible to maximize a function of these three metrics, such as their sum.

There represents the evolution, as a function of the threshold, of these three metrics, namely the precision (PRECISION), the recall (RECALL) and the recall per installation (RECALL*), obtained for the portion of the image represented in .

In this example, the threshold making it possible to maximize a function of the three aforementioned metrics is substantially equal to 0.37. The threshold thus chosen constitutes a parameter of the model during the detection step 530.

The step 540 of grouping the positive vignettes into blocks makes it possible to reconstitute the photovoltaic installations distributed over several vignettes. To achieve this, each positive thumbnail is grouped with any adjacent positive thumbnail, i.e. having a common side with the considered thumbnail, to form a block.

There represents for example the blocks B1 and B2 obtained with the positive thumbnails of the .

It can be easily understood that the smaller the thumbnails, the less the blocks will be extended. The thumbnails of FIG. 2 will for example be grouped into three blocks.

In addition, the CAMs associated with the positive vignettes are also grouped into blocks. Thus, for each photovoltaic installation detected, an aerial view and a CAM of said installation can be obtained.

The blocks obtained at the end of the grouping step 540 are then filtered during the filtering step 550 in order to eliminate any false positives.

The filtering step 550 consists in filtering the blocks formed according to three criteria: a size criterion, a color criterion and a texture criterion.

Regarding the size (or surface) criterion, and in view of the main application targeted by the applicant, namely the detection of photovoltaic installations of the solar power plant type, the algorithms have been configured to remove the blocks having a lower surface. at 20m ² for example. This size limit makes it possible to eliminate the vast majority of individual solar thermal installations (panels on private roofs used for domestic water heating).

Regarding the color criterion, any block that does not have colors similar to the usual colors of photovoltaic panels is eliminated. For example, the green channel according to RGB coding ( Red, Green, Blue ) is a good discriminator for vegetated areas.

Using a dataset of detected facilities that has been labeled by hand (true-positive or false-positive), the proportion of the green channel for each of the facilities can be visualized. False positives linked to images of forests for example can then be removed by setting a limit on the proportion of the green channel.

A colorimetric filter can then be defined by defining a function of the RGB channels of the image, for example G/(R+G+B), and observing whether there is a threshold beyond which only false -positives. The formula and the threshold are fixed and applied to all detections in order to automatically remove certain false positives by the color criterion.

Concerning the texture criterion, the blocks whose CAMs present weakly marked activations (low contrast and/or few high values of the activation) are eliminated.

There illustrates a false-positive FP and a true-positive TP on the high line and the low line respectively, with from left to right the thumbnail, the resulting CAM and the CAM histogram. Each histogram represents the number of pixels of the associated CAM (ordinate) as a function of the activation value (abscissa), the latter being a gray level between 0 and 255 as mentioned previously. It can be seen that, unlike the detected false-positive, the true-positive shows a histogram with a high number of pixels with low activation value (peak).

In addition, all the blocks whose CAM has a proportion of "gray" pixels above a certain threshold are classified as false positives. A pixel is said to be gray when its activation value is between 50 and 205.

During a real detection (true-positive), the proportion of gray pixels is very low because the associated CAM contains high values on the panels and low values everywhere else (see practical work on the ). On the contrary, the false detections (false-positives) are correlated with CAMs comprising many pixels of average activation (gray pixels).

In order to calculate the proportion of gray pixels, the distribution of pixel activation values (number of pixels per possible value from 0 to 255) is plotted, then the resulting histogram is integrated between 50 and 205. This gives a count of gray pixels as defined. The value obtained from the count is then divided by the product of the number of pixels and the sum of the activation values of the pixels of the image. The images for which a value greater than v = 3.10 ^-5 is obtained are removed.

This value of v is obtained from the classification of some of the installations detected (true-positive or false-positive): the proportion of gray pixels is evaluated for each installation detected using the method presented above, then each installation is placed on an axis according to this obtained value. The threshold value is the smallest value beyond which only false positives are obtained.

The estimation step 560 makes it possible to know with precision the coordinates and the surface of each photovoltaic installation detected.

The CAM generated for a block does not give a precise view of the pixels making up the detected installation, at least the contours of said installation appear blurred on the CAM. In order to accurately determine the coordinates and the surface of the installation, it is necessary to clearly identify the contours of the installation by an adequate analysis.

A segmentation is then applied to each visible block in order to isolate a group of pixels which corresponds to the photovoltaic installation. More precisely, the segmentation makes it possible to partition the image into a certain number of groups (portions of roads, roofs, cars, etc.). Each identified group is then classified according to whether it represents a photovoltaic panel or not from the activation of the pixels constituting said group.

This segmentation can be performed with a K-means partitioning algorithm ( K- means in English terminology) with optimized parameters.

In the groups identified by the segmentation, the activation values of the pixels are analyzed to determine if a group corresponds to a part of the photovoltaic installation (panel or part of a panel), in other words if a part of the pixels of a group shows an activation value above the threshold. If this is the case, the group concerned therefore corresponds to a part of the photovoltaic installation and the shape of said group then makes it possible to perform a binary activation of the CAM to obtain a precise shape of the part of the installation detected.

Indeed, a group of the image is identified as a photovoltaic panel if a proportion p of the pixels constituting it has an activation greater than the classification threshold. In particular, the few isolated pixels with high activation are not counted. In addition, if part of the photovoltaic installation does not appear on the CAM, it can be counted in the area calculation, the activation being done by group of pixels rather than pixel by pixel.

Therefore, this method provides better accuracy in calculating panel areas than the more basic method originally implemented in the DeepSolar model.

There represents an example of implementation of the aforementioned operations on a block resulting from the grouping step 540.

The segmentation method used is a five-dimensional K-means partitioning: channels (R,G,B) and position (X,Y). In the Python language, this method is implemented in the Scikit-image module under the name slic .

For each group identified by the algorithm, the activation values of the pixels constituting it are retrieved. For example, according to a precise adjustment of the parameters of the algorithm, a group 50 is identified as a photovoltaic panel when a proportion p = 0.31 of its pixels have an activation greater than a threshold of 0.55.

On the , pixel 51 has an activation greater than the threshold and is part of the proportion of pixels belonging to group 50 identified as a photovoltaic panel.

A binary image BA is then obtained by filling all the activated groups with 1 and the rest with 0. This image gives a faithful representation of the surface of the photovoltaic panel on the original block.

The adjustment of parameters accompanying the segmentation can also be optimized, by an optimization method for noisy or stochastic functions such as the “ simultaneous perturbation stochastic approximation ” method implemented in Python in the NoisyOpt package.

Finally, the geometric coordinates of the detected panel vertices and their surfaces are extracted from the identified shape during segmentation and binary activation.

The power determination step 570 consists in estimating the peak power of the identified photovoltaic installations.

For each installation, the peak power is determined from the area of said installation using a transfer function calculated on truth data comprising the areas and powers of the installations in the territory concerned. The truth data can come from a region similar to the region whose solar park we are trying to map.

Thus, step 570 requires the construction of an area – peak power transfer function which can then be used for future detections.

This transfer function is constructed from photovoltaic installations whose following characteristics are known: location, peak power and date of construction. The graphical representation of the powers as a function of the surfaces then makes it possible to define a strategy for constructing the transfer function.

There gives an example of such a graphical representation of the power of the power stations according to their surface, each point corresponding to a photovoltaic power station.

The strategy is based on the identification of surface/power classes and the construction of a piecewise linear transfer function.

There illustrates linear regressions to obtain the transfer function.

Initially, the plant 20 close to 20000m ² is withdrawn because it seems to present a singularity compared to the other plants. Secondly, a first linear regression (in em-dot line) is calculated for surfaces between 0 and 2000m ² , as well as a second linear regression (in broken line) for surfaces greater than 3000m ² . Finally, these two linear regressions are linearly linked between 2000m ² and 3000m ² to arrive at the transfer function represented by a continuous line.

Finally, the step 580 for eliminating the remaining false positives consists of a visual check by a human operator on all the photovoltaic installations identified in order to achieve 100% detection accuracy.

It clearly emerges from the present description that certain steps and operations of the detection method can be modified, replaced or deleted without departing from the scope of the invention.

Claims (10)

Method (500) for detecting objects (10) on a previously acquired aerial or satellite image (100), implementing a deep learning model (200) executed on a computer and comprising: a step (520) of cutting the image into a plurality of thumbnails adapted to an input format of the template; a step (530) of detecting objects (10) on the image comprising an activation calculation (531) on each thumbnail and giving a classification (532) and a CAM class activation map (533) of said thumbnail; characterized in that it comprises: a step (540) of grouping adjacent thumbnails of the same classification into blocks; a block filtering step (550); and a step (560) of estimating the geographical coordinates and the surface of each object detected from a segmentation of said blocks. Method according to claim 1, in which the image (100) is of high resolution, preferably less than 30cm. Method according to claim 1 or 2, in which the classification (532) is Boolean and results for each thumbnail from the comparison of the activation of said thumbnail with a threshold (430) chosen by an operator. A method according to claim 3, wherein the threshold (430) chosen is that maximizing a function whose variables are the precision and recall metrics calculated on a set of classification calibration data. Method according to claim 4, in which the chosen threshold (430) maximizes a function also taking as a variable a recall metric per object, defined as the percentage of objects (10) having at least one part detected on a thumbnail of a calibration image among the objects (10) visible on said image. Method according to any one of the preceding claims, in which the step (550) of filtering the blocks consists in filtering the blocks of the visible thumbnails and/or the blocks of the associated CAM activation maps according to at least one criterion among: the size, color and texture of the detected object, to eliminate false positives. A method according to any preceding claim, wherein the segmentation comprises k-means partitioning of the visible thumbnail blocks and pixel analysis of the associated CAM activation maps. Method according to any one of the preceding claims, in which the objects (10) are solar and/or photovoltaic panels which can be grouped into installations. Method according to claim 8, further comprising a step (570) of determining a nominal power for each installation detected. Computer program product characterized in that it comprises a set of program code instructions which, when executed by a processor, implement a method (500) according to one of Claims 1 to 9.