PaperlayerView exact revision on arXiv

Source-linked AI summary

A Novel CNN-LSTM-based Approach to Predict Urban Expansion

Wadii Boulila, Hamza Ghandorh, Mehshan Ahmed Khan, Fawad Ahmed, Jawad Ahmad

arXiv:2103.01695v1cs.CVcs.LG

TL;DR

Predicting urban expansion from time-series satellite images is challenging, particularly given the scale and dimensionality of remote-sensing data. The paper combines semantic segmentation with CNN-LSTM prediction and reports improved performance against commonly used algorithms across several evaluation measures. Its experiments use multi-date images from three Saudi cities.

  • Problem

    High-dimensional, large-volume satellite imagery is challenging to analyze, while land-cover prediction is often based on costly, time-consuming ground observation.

    Source: §1 · p. 2

  • Method

    The approach segments satellite images to extract urban regions and prepare training data, then couples CNN and LSTM methods to predict urban expansion.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    The proposed method performed well against commonly used algorithms in MSE, RMSE, PSNR, SSIM, and overall classification accuracy.

    Source: §8 · p. 20

  • Takeaways & Limitations

    The authors state that the approach can provide decision support for plausible future land-cover change and help identify probable future risk areas at spatiotemporal scale.

    Source: §7 · p. 19

    The increasing volume of satellite images makes applying the approach with a traditional computing framework very difficult.

    Source: §8 · p. 21

Abstract

from arXiv · show

Time-series remote sensing data offer a rich source of information that can be used in a wide range of applications, from monitoring changes in land cover to surveilling crops, coastal changes, flood risk assessment, and urban sprawl. This paper addresses the challenge of using time-series satellite images to predict urban expansion. Building upon previous work, we propose a novel two-step approach based on semantic image segmentation in order to predict urban expansion. The first step aims to extract information about urban regions at different time scales and prepare them for use in the training step. The second step combines Convolutional Neural Networks (CNN) with Long Short Term Memory (LSTM) methods in order to learn temporal features and thus predict urban expansion. In this paper, experimental results are conducted using several multi-date satellite images representing the three largest cities in Saudi Arabia, namely: Riyadh, Jeddah, and Dammam. We empirically evaluated our proposed technique, and examined its results by comparing them with state-of-the-art approaches. Following this evaluation, we determined that our results reveal improved performance for the new-coupled CNN-LSTM approach, particularly in terms of assessments based on Mean Square Error, Root Mean Square Error, Peak Signal to Noise Ratio, Structural Similarity Index, and overall classification accuracy.

1 Introduction

Time-series remote sensing provides repeated spatial information for monitoring land-cover change, but high-resolution satellite imagery is challenging to analyze at scale. The paper proposes a two-step approach that segments urban regions before using coupled CNN and LSTM methods to predict expansion.

1 Introduction

  • Satellite revisits provide high-spatial-resolution images across time, enabling characterization of land cover at key dates.The paper describes time-series remote sensing as a cost-effective source of spatial information and temporal patterns.Source: §1 · p. 1
  • Remote sensing offers a practical alternative to costly, time-consuming on-the-ground data gathering for predicting land-cover maps.The paper notes that remote sensing data are used by governments, decision-makers, and planners.Source: §1 · p. 2
  • High dimensionality and large image volumes challenge remote-sensing analysis, while traditional computer-vision algorithms explore only a small portion of satellite data.The paper notes that satellite sensors capture data at high spatial, temporal, and spectral resolutions.Source: §1 · p. 2
  • Deep-learning models can automatically extract task-specific features and learn temporal and long-term dependencies from multidimensional time-series data.The paper notes that model accuracy depends on network architecture and training data.Source: §1 · p. 2
  • The approach first segments multi-date satellite images to extract urban regions and prepare training data, then couples CNN and LSTM methods to predict urban maps for upcoming dates.The prediction step aims to learn temporal features.Source: §1 · p. 2

2 Literature review

Prior remote-sensing change prediction spans statistical, hybrid, Markov, and AI-based methods, with examples across land cover, forests, crops, and other applications. The paper identifies relatively little work on urban-expansion prediction and proposes unsupervised segmentation alongside ConvLSTM for that task.

2 Literature review

  • Earlier remote-sensing land-cover, forest, and vegetation change studies used statistical, hybrid, and Markov forecasting methods, alongside increasingly common AI-based models.The review notes AI methods' ease of implementation and flexibility with large data volumes.Source: §2 · p. 3
  • CA-Markov was used to simulate mangrove land-cover changes, while logistic regression predicted forest and non-forest areas with 87% probability.The cited mangrove study projected 21ha changing to open water and 28ha expanding landward by 2025.Source: §2 · p. 3
  • Crop-yield prediction combined histogram-based dimensionality reduction, CNN, LSTM, and a Gaussian Process layer, while Deep-STEP and SARDINE addressed satellite-data prediction.The review reports that Deep-STEP produced fewer errors than MLP, DSN, and NARNET, and that SARDINE saved computational time on large satellite images.Source: §2 · p. 3
  • Other reviewed applications used supervised self-organizing maps for wheat-yield variation and graph convolution with LSTM for atmospheric PM2.5 prediction.SKN had the best accuracy among the three wheat models; GC-LSTM performed better than RNN, BP neural network, or convolution LSTM.Source: §2 · p. 4
  • The authors identify few studies predicting urban expansion and describe accuracy limits and user intervention in existing approaches, positioning their work as the first ConvLSTM model for urban growth to their knowledge.They also propose semi-ground-truth extraction with unsupervised segmentation to separate target areas for prediction.Source: §2 · p. 4

3 Preliminaries

The preliminaries describe unsupervised segmentation for extracting image regions, ConvLSTM for retaining spatial structure while learning from sequences, and batch normalization for standardizing layer inputs.

3.1 Unsupervised image segmentation

  • Unsupervised segmentation partitions an image into meaningful regions without prior label knowledge, avoiding the need for manually labeled data.Kim et al.'s approach assigns pixel cluster labels through learned feature-extraction and mapping functions without human supervision.Source: §3.1 · pp. 4–5
  • The segmentation network uses convolution, ReLU, and batch normalization to compute feature maps, then selects cluster labels by argmax from normalized classifier responses.Its loss combines feature similarity and spatial continuity; μ balances these constraints, and the filters are updated by backpropagation.Source: §3.1 · p. 5

3.2 Convolution LSTM

  • ConvLSTM replaces fully connected operators with convolutions, allowing sequential-image processing that preserves spatial feature-map size while learning global spatiotemporal information.The paper contrasts this with vectorized LSTM inputs, which lose spatial correlation.Source: §3.2 · pp. 5–6
  • In ConvLSTM, cell and hidden states are updated through input, forget, and output gates, with convolution kernels and bias terms.The equations use convolution and Hadamard products; the sigmoid activation modulates gate outputs between 0 and 1.Source: §3.2 · p. 6

3.3 Batch Normalization (BN)

  • Batch normalization standardizes each layer's input distribution and preserves activations with mean 0 and variance 1 to help accelerate training.Its transformation includes learnable parameters γ and β, and uses ε to avoid numerical instability.Source: §3.3 · p. 6

4 Dataset

The study evaluates urban expansion in Riyadh, Jeddah, and Dammam using multi-date satellite imagery. The dataset uses four images per city time series, expert-supported urban ground truth, and corrected and registered imagery.

4 Dataset

  • The dataset covers Riyadh, Jeddah, and Dammam, selected as large cities with notable urban growth and used to evaluate and validate the approach.The paper describes Riyadh as the capital and financial hub, Jeddah as a seaport and second-largest city, and Dammam as the Eastern Province capital and oil-industry center.Source: §4 · p. 6
  • The study characterizes rapid urban expansion in the three cities and attributes it mainly to socioeconomic growth associated with rising oil revenues.The paper presents the cities as a relevant case for studying regional land-cover change.Source: §4 · p. 7
  • Experiments use three-part time series of satellite imagery captured by the French SPOT satellite.Table 1 is described as listing region, acquisition date, resolution, image vertex locations, and sensor type.Source: §4 · p. 7, Table 1 · p. 7
  • The three city time series each contain four satellite images with a two-year temporal frequency, and the images were radiometrically and geometrically corrected and spatially registered to sub-pixel accuracy.The six land-cover types identified by domain experts are water, urban, soil, vegetation, mountain, and road.Source: §4 · p. 7
  • Urban ground truth was generated with domain experts using unsupervised segmentation alongside Google Maps, extracting clusters of buildings and houses.The study says about 40% of land cover across the three regions changed to urban areas between 2006 and 2013.Source: §4 · p. 7

5 Proposed approach

The approach segments satellite images to prepare urban masks, then uses ConvLSTM to predict urban expansion from image sequences. The workflow combines image cleanup with model training on successive dates.

5.1 Image segmentation

  • Satellite imagery is segmented in non-overlapping 1024 × 1024 blocks, then non-urban classes and residual objects are removed to prepare urban masks for training.The target is the blue urban class; residual objects are removed manually with ENVI's ROI tool after image-processing operations.Source: §5.1 · pp. 8–9, Figure 2 · p. 9
  • Urban-expansion prediction uses four ConvLSTM layers with 3 × 3 kernels and 40 filters each, three batch-normalization layers, and an output convolution.Source: §5.1 · p. 10
  • Because the remote-sensing images are too high-resolution to process directly, they are divided into non-overlapping 256 × 256 blocks with pixel values normalized to [0,1].Source: §5.1 · p. 11
  • Training pairs 2015 satellite images as input with 2017 images as output, while validation pairs 2017 inputs with 2019 outputs.Source: §5.1 · p. 11
  • The model is trained with randomly initialized weights and batch size 10, minimizing categorical cross-entropy using Adam.The reported Adam learning rate is 0.001, with β1 = 0.9 and β2 = 0.999.Source: §5.1 · pp. 10, 12

6 Experimental results

ConvLSTM is evaluated against Pix2Pix GAN and Dual GAN on validation blocks and a 2019 satellite image. It performs better across the reported mean metrics, although some block-level metric values tie.

6 Experimental results

  • ConvLSTM training reached 92.19% validation accuracy with categorical cross-entropy loss of 0.242 after 32 epochs.Training used categorical cross-entropy, batch size 10, and Adam; the authors report no further accuracy increase or loss decrease after epoch 32.Source: §6 · p. 12, Figure 7 · p. 13
  • Pix2Pix trained fastest at 3.16 hours, followed by ConvLSTM at 4.17 hours and Dual GAN at 6.39 hours.Pix2Pix ran for 35 epochs and Dual GAN for 37 epochs until each reached its minimum loss threshold.Source: §6 · p. 13, Table 3 · p. 13
  • Most validation-block error values are lower for ConvLSTM, and most PSNR and SSIM values are higher, though some values tie across models.Figures 10–13 show the validation-block metric values; lower MSE/RMSE and higher PSNR/SSIM indicate better predictions.Source: §6 · p. 14, Figure 10 · p. 15, Figure 11 · p. 15
  • Across mean MSE, RMSE, PSNR, and SSIM, ConvLSTM produces better results than Pix2Pix and Dual GAN.The comparison is reported in Table 4.Source: Table 4 · p. 16, §6 · p. 16
  • For the January 11, 2019 image, Figure 15’s confusion matrix uses ground-truth urban/non-urban classes as rows and predicted classes as columns.The accompanying text reports overall accuracy of 90.30% for Pix2Pix GAN and 88.88% for Dual GAN.Source: §6 · p. 18, Figure 15 · p. 19

7 Discussions

The discussion presents the CNN–LSTM approach as a tool for studying urban expansion in Riyadh, Jeddah, and Dammam. It links rapid city growth to planning pressures and describes distinct growth patterns across the cities.

7 Discussions

  • The study applies a fused CNN–LSTM approach to a multi-date satellite dataset for Saudi Arabia’s three largest cities.The authors frame the results as potential decision support for future land-cover change and urban-expansion planning.Source: §7 · p. 19
  • Saudi Arabia’s rapid urban growth is attributed mainly to economic growth driven by oil revenue.The discussion notes that more than 80% of the country’s population lives in urban areas and that the study focuses on Riyadh, Jeddah, and Dammam.Source: §7 · p. 20
  • Riyadh’s expansion spread outward from the center; Jeddah grew more to the north and east than south, while Dammam expanded west along the Arabian Gulf.In most cases, expansion occurred through existing urban land cover rather than separately from it.Source: §7 · p. 20
  • Urban areas expanded faster than populations in all three cities from 2013 to 2019: over threefold in Riyadh, over twofold in Jeddah, and 400% in Dammam.The reported populations rose from around 5.6 to 7 million in Riyadh, 3.9 to 4.5 million in Jeddah, and 0.7 to 1.2 million in Dammam.Source: §7 · p. 20
  • The authors associate rapid urbanization in the three regions with challenges for regional planning and decision-makers.They call for compact, high-density, mixed-use development and bylaws that accommodate more people in the same area.Source: §7 · p. 20

8 Conclusion

The conclusion describes a two-step approach combining semantic segmentation with urban-expansion prediction and reports strong performance across several metrics. It presents the method as useful for sustainable urban planning while identifying large remote-sensing datasets as a future challenge.

8 Conclusion

  • The proposed two-step method combines semantic image segmentation with urban-expansion prediction and performs well on MSE, RMSE, PSNR, SSIM, and overall classification accuracy.The study evaluates the approach on the three largest cities in Saudi Arabia.Source: §8 · p. 20
  • The authors consider the approach a useful tool for sustainable urban planning.This conclusion is based on the reported results.Source: §8 · p. 20
  • Applying the approach to large remote-sensing datasets remains a future direction because growing image volumes are difficult to handle with traditional computing frameworks.The authors also suggest investigating security for large satellite images shared across servers.Source: §8 · p. 21
Loading 2103.01695v1…