Work

Deep learning combined with focus on important feature strategy for the detection of Dendrolimus infestation area

Year: 2025

Type: article

Abstract: Recent methods using remote sensing satellite images for detecting Dendrolimus infestation area (DIA) struggle to fully utilize image information, leading to issues like low accuracy, recall, and adaptability. To address these issues, we propose the idea of using enhanced focus on important feature strategies and deep learning. We constructed a combination network of Noteworthy Various Feature Capture methods and Fully Connected Network (NVCF). NVCF incorporates a multi-scale sampling module (MSSM) and featuring channel and spatial attention module (SAM). We have developed improved channel attention module (IPCAM) and integrated it into the NVCF along with the SAM, enabling it to learn more efficiently from the channel and location information. IPCAM, SAM, multi-scale input module (MCM), and dilated convolution (AC) modules together form the MSSM. The MCM and AC capture multi-scale features and extend the receptive field of each network layer. We found that reflectance mutations in the DIA occur within the 740–1300 nm range and further constructed the Normalized Near-Infrared Dendrolimus Index (NNDI). The NNDI, vegetation indices (Veg) and non-vegetation indices (Non-Veg), were used together in the NVCF model training. The model was evaluated using a dataset from a typical region on the northern slope of Changbai Mountain in China. NVCF model achieves a recall of 0.95 and a Kappa of 0.97, surpassing the performance of PWDNet + BM, PWDNet, and D-SCNet. We conducted experiments on actual application scenarios, and the results showed the reliability and good adaptability of DIA detection of the NVCF. All the above results indicate that our method can provide technical support for the detection and control of DIA.

Source: Computers and Electronics in Agriculture

DOI: 10.1016/j.compag.2025.110780