Solid waste mapping based on very high resolution remote sensing imagery and a novel deep learning approach

Figures & data

Figure 1. Study areas.

Figure 2. Samples of solid waste sites. (a) Langfang city in China; (b) Faridabad city in India; (c) Tezoyuca city in Mexico.

Table 1. Classification scheme.

Study area	Image size	Category	Training	Testing	Total
Langfang, China	34048 × 34720	SW	552	138	690
		non-SW	552	138	690
Faridabad, India	16095 × 16582	SW	520	130	650
		non-SW	520	130	650
Tezoyuca, Mexico	14905 × 14700	SW	400	100	500
		non-SW	400	100	500
		Total	2944	736	3680

Note. SW stands for solid waste sites and non-SW represents non-solid waste sites.

Figure 3. Overall structure of the proposed SW-Net.

Figure 4. CNN stream.

Figure 5. Transformer stream.

Figure 6. Gated fusion module.

Figure 7. Boundary delineation of solid waste sites. (a) remote sensing image patch; (b) image patch overlayed with CAM; and (c) boundaries of solid waste sites generated by thresholding the CAM image.

Figure 8. Mapping result of solid waste sites for (a) Langfang; (b) Faridabad; (c) Tezoyuca.

Figure 9. Confusion matrix of each study area. Notes. 0 represents non-solid waste sites and 1 denotes solid waste sites.

Figure 10. Examples of several predicted image patches. (a) solid waste sites predicted as non-solid waste sites; (b) non-solid waste sites predicted as solid waste sites.

Figure 11. Exposure risk map for (a) Langfang; (b) Faridabad; (c) Tezoyuca.

Table 2. Accuracy for CNN-only, Transformer-only and the proposed model.

Method\Study area	OA (%) (A)	OA (%) (B)	OA (%) (C)	Average (%)
CNN-only	86.96	88.85	73.00	83.83
Transformer-only	84.78	85.77	71.50	81.52
Feature-stacking	88.77	88.08	74.50	84.64
Proposed	94.20	94.62	80.50	90.62

Note. A, B and C stand for Langfang, Faridabad and Tezoyuca, respectively.

Table 3. Comparison with other deep learning models.

Method\Study area	OA (%) (A)	OA (%) (B)	OA (%) (C)	Average (%)
VGG-16	82.25	85.77	70.00	80.16
ResNet-50	82.61	88.46	72.50	81.93
EfficientNet-B0	81.52	77.31	60.50	74.32
DenseNet-121	89.49	88.85	76.00	85.59
Proposed	94.20	94.62	80.50	90.62

Note. A, B and C stand for Langfang, Faridabad and Tezoyuca, respectively.

Table 4. Comparison with other solid waste classification methods.

Approach	Data	Muti-Region	Model	Accuracy
This work	VHR image	Yes	CNN-Transformer	90.62%
Silvestri and Omri (2008)	VHR image	No	NDVI based	–
Lin et al. (2022)	Ground image	No	VGG-like	76.8%
Bao et al. (2018)	UAV image	No	Thresholding	98.6%
Fallati et al. (2019)	UAV image	Yes	CNN	94%
Sun et al. (2021)	VHR image	Yes	SRAF-Net	83.17%

Silvestri S, Omri M. 2008. A method for the remote sensing identification of uncontrolled landfills: formulation and validation. Int J Remote Sens. 29(4):975–989.

 Web of Science ®Google Scholar

Lin K, Zhao Y, Kuo JH, Deng H, Cui F, Zhang Z, Zhang M, Zhao C, Gao X, Zhou T, et al. 2022. Toward smarter management and recovery of municipal solid waste: a critical review on deep learning approaches. J Clean Prod. 346:130943.

 Web of Science ®Google Scholar

Bao Z, Sha J, Li X, Hanchiso T, Shifaw E. 2018. Monitoring of beach litter by automatic interpretation of unmanned aerial vehicle images using the segmentation threshold method. Mar Pollut Bull. 137:388–398. 10.1016/j.marpolbul.2018.08.009. 30503448

 PubMed Web of Science ®Google Scholar

Fallati L, Polidori A, Salvatore C, Saponari L, Savini A, Galli P. 2019. Anthropogenic marine debris assessment with unmanned aerial vehicle imagery and deep learning: a case study along the beaches of the Republic of Maldives. Sci Total Environ. 693:133581.

 PubMed Web of Science ®Google Scholar

Sun X, Liu Y, Yan Z, Wang P, Diao W, Fu K. 2021. SRAF-Net: shape robust anchor-free network for garbage dumps in remote sensing imagery. IEEE Trans Geosci Remote Sensing. 59(7):6154–6168.

 Web of Science ®Google Scholar

Data availability statement

The data & code that support the findings of this study are openly available at [https://github.com/MrSuperNiu/Remote-Sensing-for-Solid-Waste-mapping].