[Paper Review] Greenhouse Segmentation on High-Resolution Optical Satellite Imagery using Deep Learning Techniques

Orkhan Baghirli, Imran Ibrahimli+1 more|arXiv (Cornell University)|Jul 22, 2020

Remote Sensing and LiDAR Applications23 references4 citations

TL;DR

This paper proposes a U-Net-based deep learning framework for high-accuracy greenhouse segmentation in 1.5-meter resolution optical satellite imagery from Azerbaijan. By integrating dilated convolutions, skip connections, bilinear interpolation, and a weighted binary cross-entropy + dice loss, the best model achieves a 93.29% F1 score, outperforming the baseline U-Net by 4.48%.

ABSTRACT

Greenhouse segmentation has pivotal importance for climate-smart agricultural land-use planning. Deep learning-based approaches provide state-of-the-art performance in natural image segmentation. However, semantic segmentation on high-resolution optical satellite imagery is a challenging task because of the complex environment. In this paper, a sound methodology is proposed for pixel-wise classification on images acquired by the Azersky (SPOT-7) optical satellite. In particular, customized variations of U-Net-like architectures are employed to identify greenhouses. Two models are proposed which uniquely incorporate dilated convolutions and skip connections, and the results are compared to that of the baseline U-Net model. The dataset used consists of pan-sharpened orthorectified Azersky images (red, green, blue,and near infrared channels) with 1.5-meter resolution and annotation masks, collected from 15 regions in Azerbaijan where the greenhouses are densely congested. The images cover the cumulative area of 1008 $km^2$ and annotation masks contain 47559 polygons in total. The $F_1, Kappa, AUC$, and $IOU$ scores are used for performance evaluation. It is observed that the use of the deconvolutional layers alone throughout the expansive path does not yield satisfactory results; therefore, they are either replaced or coupled with bilinear interpolation. All models benefit from the hard example mining (HEM) strategy. It is also reported that the best accuracy of $93.29\%$ ($F_1\,score$) is recorded when the weighted binary cross-entropy loss is coupled with the dice loss. Experimental results showed that both of the proposed models outperformed the baseline U-Net architecture such that the best model proposed scored $4.48\%$ higher in comparison to the baseline architecture.

Motivation & Objective

• To develop a robust semantic segmentation pipeline for identifying greenhouses in high-resolution optical satellite imagery.
• To address challenges such as scale variation, irregular shapes, shadows, and atmospheric distortions in greenhouse detection.
• To improve segmentation accuracy over baseline U-Net by incorporating architectural enhancements like dilated convolutions and bilinear upsampling.
• To evaluate the impact of loss functions, data augmentation, and hard example mining on model generalization and performance.
• To create a computationally feasible and deployable solution for climate-smart agricultural monitoring using satellite data.

Proposed method

• Utilized pan-sharpened, orthorectified Azersky (SPOT-7) satellite images with 1.5m resolution, including RGB and near-infrared bands.
• Preprocessed images by splitting into 64×64 tiles with 32-pixel overlap and retained only tiles with positive class rates ≥0.1 to address class imbalance.
• Enhanced features by computing NDVI and texture channels and applied data augmentation techniques including mixup, rotation, flipping, and brightness/gamma adjustments.
• Designed two novel U-Net variants: Model A and Model B, incorporating dilated convolutions and skip connections to capture multi-scale context.
• Replaced or combined transposed convolutions with bilinear interpolation in the decoder path to reduce checkerboard artifacts.
• Optimized training using hard example mining (HEM), adaptive learning rate scheduling, and a hybrid loss combining weighted binary cross-entropy and dice loss.

Experimental results

Research questions

• RQ1 How do dilated convolutions and skip connections improve greenhouse segmentation in high-resolution satellite imagery?
• RQ2 What is the impact of replacing transposed convolutions with bilinear interpolation on segmentation quality and artifact reduction?
• RQ3 How does the combination of weighted binary cross-entropy and dice loss affect F1 and AUC scores compared to standard loss functions?
• RQ4 To what extent do data augmentation and hard example mining improve model robustness under challenging conditions like shadows and haze?
• RQ5 How do architectural variations between the baseline U-Net and the proposed models affect performance on irregular, small, or shadowed greenhouses?

Key findings

• The best-performing model (Model B) achieved a 93.29% F1 score, representing a 4.48% improvement over the baseline U-Net.
• Model B demonstrated superior performance in complex scenarios, such as shadowed or irregularly shaped greenhouses, where it correctly segmented all objects without gaps.
• The use of bilinear interpolation instead of transposed convolutions significantly reduced checkerboard artifacts and improved visual quality of segmentation masks.
• The hybrid loss function (weighted binary cross-entropy + dice loss) yielded the highest AUC (95.91%) and F1 score, indicating strong generalization and robustness.
• Hard example mining and data augmentation significantly improved model performance on difficult cases, especially under poor atmospheric conditions.
• Model A showed intermediate performance, outperforming the baseline but underperforming Model B, particularly in detecting small or partially obscured greenhouses.