An ABM Approach to Addressing the Opioid Crisis: How can we predict the effects of public health intervention on a unique community of pain patients?
Detecting Solar Panels From Satellite Imagery
Convolutional Neural Networks handle pixels in context making them an ideal choice for image recognition.
Contributors to this project include Viggy Kumaresan, Azucena Morales, Yifei Wang, and Sicong Zhao.
Why Bother With Solar Panel Detection?
Solar power currently accounts for 1% of the world’s electricity generation. In fact, estimates of solar energy production predict a potential 65-fold growth by 2050, eventually making solar power one of the largest sources of energy across the globe [1]. Solar photovoltaic, or solar PV, power installed on top of rooftops is estimated to make up 30% of this energy generation. In recent years, solar PV power has already begun playing an increasingly large role in US electricity generation. From 2008 through 2017, there was a 39-fold growth in annual solar generation representing an increase of 75,123 GWh [2].
As consumer solar PV adoption increases, so does interest in solar consumption habits. Policymakers, for example, are reliant on accurate measures of the saturation of solar PV to inform structuring of tax incentives programs [3, 4]. Solar manufacturers also rely on detailed solar PV market insights to acquire new customers and customize sales strategies [5, 6].
Traditional data sources such as consumer surveys and market research, are costly and time-consuming to collect, and ultimately can often only give a partial or biased view of the market. Satellite imagery, on the other hand, allows us to see overhead views of households all over the country and does not rely on self-reported data. The use of aerial imagery to for solar PV detection may result in more consistent and cost-effective assessment of solar adoption.
Previous Research
Previous studies have explored the use of machine learning for satellite image classification with high rates of accuracy of detection with a low rate of incorrect classification [7, 8]. Specifically, random forest classification and CNN approaches have been shown to accurately measure size, shape and capacity of solar PV from satellite imagery [9, 10, 11]. However, to implement approaches like these, as with any supervised learning approach, a training set of data with labelled classes is required in order to both model the data as well as evaluate performance. Unfortunately, the number of available labelled datasets containing satellite imagery of solar PV is limited. An example of such a dataset is the Distributed Solar Photovoltaic Array Location and Extent Dataset containing geospatial coordinates and border vertices for over 19,000 solar panels across 601 high-resolution images from four cities in California [12].
Data Sources
Images Containing Solar PV
Images Not Containing Solar PV
The primary dataset contains 1,500 satellite images in TIFF format, with each image consisting of 101 pixels x 101 pixels and three color channels (Red, Green, Blue). In total, each image is represented as 101 x 101 x 3 array, which is a total of 30,603 numerical entries with values ranging from 0 and 255.
Each image in the dataset has been labeled as one of two classes, either containing a solar PV, or not containing a solar PV. The labels were created by human observers who visually scanned the imagery and annotated all of the visible solar PV. There are 555 images that contains a solar PV and 995 images without a solar PV, which means that the classes are not entirely balanced. An additional 558 labeled images are housed on Kaggle.com. These images were used as a test set during the model selection process.
Looking at the images themselves, many instances of solar panels tend to be rectangle with sharp angles and edges. That said, the full images containing solar PV do not always share a standard form. Images come in a variety of sizes and colors and the solar panels are not always at the center of the image. Moreover, the surrounding landscape in images of both classes is also far from standardized. Examples of rooftops, pavement, grass, and residential pools can be seen in both classes. In order for any model to be successful, this lack of standardization needs to be addressed. Additionally, a model should predict the a consistent class regardless of orientation of each image. In other words, rotation and scaling of images should be tolerated in the final model.
Methods
Analysis Workflow
Train-Test Split
A 80-20 training-test split was used. When partitioning the data, a ratio of classes proportional to the original dataset in both training and validation sets.
Data Preprocessing
Data preprocessing took three general steps. In the first step of preprocessing, RGB values were rescaled to be between 0 and 1 in order to avoid high values compared to a typical learning rate [13].
During the second step of preprocessing, random modifications were applied to images including horizontal flips, image shears, and zooms. By increasing the diversity of images in the training set we can ideally to reduce overfit and ultimately increase out-of-sample accuracy [13, 14, 15].
Finally, luminance and gradient features were added to each image. These features were selected based on properties we know to be true of solar panels, they absorb light and they are angular in shape. By adding these two features, we hoped to increase prediction accuracy [16].
Convolutional Neural Networks (CNN)
CNNs are a type of neural network for processing data that has a grid-like topology. CNNs are widely known in computer vision for being successful for tasks such as classifying images, clustering images, and object recognition [17]. At they’re core, CNNs use the convolution operation, instead of matrix multiplication in at least one of their layers. Similar to other neural networks, they are conformed by a sequence of layers. The layers of CNN have neurons arranged in three dimensions: width, height and depth. There are many different types of CNN architectures, but in essence, CNNs handle pixels in context to their surroundings making them an ideal choice for image recognition. [18, 19]
Results
Compared to a variety of modeling approaches, a CNN approached proved to be the strongest. KNN classification and gaussian naive bayes modeling performed comparably to each other with AUCs of 0.639 and 0.643 respectively. Logistic regression was able to achieve a surprising 0.785 AUC. Random forest modeling performed even stronger with and AUC of 0.824. Finally, CNN resulted in validation accuracy of 0.957. Although AUC and validation accuracy are not measured on the same scale, the results suggest the CNN model shows a meaningful improvement in accuracy.
Model Accuracy
| Model | Metric | Validation Set Performance |
|---|---|---|
| Chance | AUC | 0.500 |
| KNN | AUC | 0.639 |
| Gaussian Naive Bayes | AUC | 0.643 |
| Logistic Regression | AUC | 0.785 |
| Random Forest | AUC | 0.824 |
| CNN | Validation Accuracy | 0.957 |
The addition of a gradient channel increased the accuracy of many models techniques.
Logistic Regression Performance By Proprocessing Method
Random Forest Performance By Proprocessing Method
Final Thoughts
These results align with previous research suggesting that approaches such as logistic regression, k-nearest neighbors, and gaussian naive bayes may not be sufficient for identification with high rates of accuracy. That said, by adding features such as image gradients, performance can be improved even when using simple models such as logistic regression.
Moreover, these results suggest a CNN approach produces the highest accuracy of detection. Instead of handling each pixel as features individually, CNN essentially looks at pixels in context. This type of approach is ideal when considering image recognition analysis because pixels only have meaning in relation to surrounding pixels. CNN or similar neural network approaches appear to be a feasible and strong choice for similar image detection analyses.
In comparison to previous research on supervised learning and solar PV detection, this analysis was rather small. Our training and testing dataset was only half the size of similar studies [9]. By increasing the size of the training data, model accuracy may be improved. Also, additional approaches could yield increased accuracy, such as transferring learning from pre-built models. Future research should also examine the use of supervised learning techniques to not only identify the existence or absence of solar PV, but also to estimate the energy consumption associated with each installation. This could be achieved by integrating other disparate data sources, such as geographic data and energy usage patterns. Ultimately, increasingly accurate information around both consumer solar adoption and solar energy consumption would aid policymakers in the structuring of new green policies and incentive programs.
References
[1] Energy Transition Outlook 2018. DNV GL.
[2] "Renewables on the Rise 2018 - Environment America.”
[3] Matasci, S. "Solar tax credit". EnergySage. (6 Jan. 2019)
[4] Solangi, K. H., et al. "A review on global solar energy policy." Renewable and sustainable energy reviews 15.4 (2011): 2149-2163.
[5] Janes, M., 2014. "Predictive models help determine which consumers buy solar equipment and why". Phys.org. (6 Jan. 2019)
[6] Rai, Varun, D. Cale Reeves, and Robert Margolis. "Overcoming barriers and uncertainties in the adoption of residential solar PV." Renewable Energy 89 (2016): 498-505.
[7] Mnih, Volodymyr, and Geoffrey E. Hinton. "Learning to detect roads in high-resolution aerial images." European Conference on Computer Vision. Springer, Berlin, Heidelberg, 2010.
[8] Kubat, Miroslav, Robert C. Holte, and Stan Matwin. "Machine learning for the detection of oil spills in satellite radar images." Machine learning 30.2-3 (1998): 195-215.
[9] Malof, Jordan M., et al. "Automatic detection of solar photovoltaic arrays in high resolution aerial imagery." Applied energy 183 (2016): 229-240.
[10] Malof, Jordan M., Leslie M. Collins, and Kyle Bradbury. "A deep convolutional neural network, with pre-training, for solar photovoltaic array detection in aerial imagery." 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). IEEE, 2017.
[11] Golovko, Vladimir, et al. "Convolutional neural network based solar photovoltaic panel detection in satellite photos." 2017 9th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS). Vol. 1. IEEE, 2017.
[12] Bradbury, Kyle, et al. "Distributed solar photovoltaic array location and extent dataset for remote sensing object identification." Scientific data 3 (2016): 160106.
[13] Chollet, F. "Building powerful image classification models using very little data”. The Keras Blog. (June 5, 2016).
[14] Venkatesh, T. “Simple Image Classification using Convolutional Neural Network — Deep Learning in python”
[15] Lewinson, E. “Mario vs. Wario: Image Classification in Python.” (24 Jul 2018).
[16] Stokes, M., Anderson, M., Chandrasekar, S., & Motta, R. “A Standard Default Color Space for the Internet.” (1996, November 5).
[17] Karpathy, A. “CS231n Convolutional Neural Networks for Visual Recognition.” (2018).
[18] Goodfellow, Ian, Bengio Yoshua, and Courville, Aaron, 2016. “CS231n Convolutional Neural Networks for Visual Recognition.” (2018).
[19] Skymind “A Beginner's Guide to Convolutional Neural Networks (CNNs).”
[20] Prabhu, R. "Understanding of Convolutional Neural Network (CNN)." (3 Mar. 2018).